1 //===- InlineCoast.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 //===----------------------------------------------------------------------===//
15 #include "llvm/Transforms/Utils/InlineCost.h"
16 #include "llvm/Support/CallSite.h"
17 #include "llvm/CallingConv.h"
18 #include "llvm/IntrinsicInst.h"
22 // CountCodeReductionForConstant - Figure out an approximation for how many
23 // instructions will be constant folded if the specified value is constant.
25 unsigned InlineCostAnalyzer::FunctionInfo::
26 CountCodeReductionForConstant(Value *V) {
27 unsigned Reduction = 0;
28 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI)
29 if (isa<BranchInst>(*UI))
30 Reduction += 40; // Eliminating a conditional branch is a big win
31 else if (SwitchInst *SI = dyn_cast<SwitchInst>(*UI))
32 // Eliminating a switch is a big win, proportional to the number of edges
34 Reduction += (SI->getNumSuccessors()-1) * 40;
35 else if (CallInst *CI = dyn_cast<CallInst>(*UI)) {
36 // Turning an indirect call into a direct call is a BIG win
37 Reduction += CI->getCalledValue() == V ? 500 : 0;
38 } else if (InvokeInst *II = dyn_cast<InvokeInst>(*UI)) {
39 // Turning an indirect call into a direct call is a BIG win
40 Reduction += II->getCalledValue() == V ? 500 : 0;
42 // Figure out if this instruction will be removed due to simple constant
44 Instruction &Inst = cast<Instruction>(**UI);
45 bool AllOperandsConstant = true;
46 for (unsigned i = 0, e = Inst.getNumOperands(); i != e; ++i)
47 if (!isa<Constant>(Inst.getOperand(i)) && Inst.getOperand(i) != V) {
48 AllOperandsConstant = false;
52 if (AllOperandsConstant) {
53 // We will get to remove this instruction...
56 // And any other instructions that use it which become constants
58 Reduction += CountCodeReductionForConstant(&Inst);
65 // CountCodeReductionForAlloca - Figure out an approximation of how much smaller
66 // the function will be if it is inlined into a context where an argument
69 unsigned InlineCostAnalyzer::FunctionInfo::
70 CountCodeReductionForAlloca(Value *V) {
71 if (!isa<PointerType>(V->getType())) return 0; // Not a pointer
72 unsigned Reduction = 0;
73 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){
74 Instruction *I = cast<Instruction>(*UI);
75 if (isa<LoadInst>(I) || isa<StoreInst>(I))
77 else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
78 // If the GEP has variable indices, we won't be able to do much with it.
79 for (Instruction::op_iterator I = GEP->op_begin()+1, E = GEP->op_end();
81 if (!isa<Constant>(*I)) return 0;
82 Reduction += CountCodeReductionForAlloca(GEP)+15;
84 // If there is some other strange instruction, we're not going to be able
85 // to do much if we inline this.
93 /// analyzeFunction - Fill in the current structure with information gleaned
94 /// from the specified function.
95 void InlineCostAnalyzer::FunctionInfo::analyzeFunction(Function *F) {
96 unsigned NumInsts = 0, NumBlocks = 0, NumVectorInsts = 0;
98 // Look at the size of the callee. Each basic block counts as 20 units, and
99 // each instruction counts as 5.
100 for (Function::const_iterator BB = F->begin(), E = F->end(); BB != E; ++BB) {
101 for (BasicBlock::const_iterator II = BB->begin(), E = BB->end();
103 if (isa<PHINode>(II)) continue; // PHI nodes don't count.
105 // Special handling for calls.
106 if (isa<CallInst>(II) || isa<InvokeInst>(II)) {
107 if (isa<DbgInfoIntrinsic>(II))
108 continue; // Debug intrinsics don't count as size.
110 CallSite CS = CallSite::get(const_cast<Instruction*>(&*II));
112 // If this function contains a call to setjmp or _setjmp, never inline
113 // it. This is a hack because we depend on the user marking their local
114 // variables as volatile if they are live across a setjmp call, and they
115 // probably won't do this in callers.
116 if (Function *F = CS.getCalledFunction())
117 if (F->isDeclaration() &&
118 (F->isName("setjmp") || F->isName("_setjmp"))) {
123 // Calls often compile into many machine instructions. Bump up their
124 // cost to reflect this.
125 if (!isa<IntrinsicInst>(II))
129 if (const AllocaInst *AI = dyn_cast<AllocaInst>(II)) {
130 if (!AI->isStaticAlloca())
131 this->usesDynamicAlloca = true;
134 if (isa<ExtractElementInst>(II) || isa<VectorType>(II->getType()))
137 // Noop casts, including ptr <-> int, don't count.
138 if (const CastInst *CI = dyn_cast<CastInst>(II)) {
139 if (CI->isLosslessCast() || isa<IntToPtrInst>(CI) ||
140 isa<PtrToIntInst>(CI))
142 } else if (const GetElementPtrInst *GEPI =
143 dyn_cast<GetElementPtrInst>(II)) {
144 // If a GEP has all constant indices, it will probably be folded with
146 bool AllConstant = true;
147 for (unsigned i = 1, e = GEPI->getNumOperands(); i != e; ++i)
148 if (!isa<ConstantInt>(GEPI->getOperand(i))) {
152 if (AllConstant) continue;
161 this->NumBlocks = NumBlocks;
162 this->NumInsts = NumInsts;
163 this->NumVectorInsts = NumVectorInsts;
165 // Check out all of the arguments to the function, figuring out how much
166 // code can be eliminated if one of the arguments is a constant.
167 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E; ++I)
168 ArgumentWeights.push_back(ArgInfo(CountCodeReductionForConstant(I),
169 CountCodeReductionForAlloca(I)));
174 // getInlineCost - The heuristic used to determine if we should inline the
175 // function call or not.
177 InlineCost InlineCostAnalyzer::getInlineCost(CallSite CS,
178 SmallPtrSet<const Function *, 16> &NeverInline) {
179 Instruction *TheCall = CS.getInstruction();
180 Function *Callee = CS.getCalledFunction();
181 Function *Caller = TheCall->getParent()->getParent();
183 // Don't inline a directly recursive call.
184 if (Caller == Callee ||
185 // Don't inline functions which can be redefined at link-time to mean
187 // FIXME: We allow link-once linkage since in practice all versions of
188 // the function have the same body (C++ ODR) - but the LLVM definition
189 // of LinkOnceLinkage doesn't require this.
190 (Callee->mayBeOverridden() && !Callee->hasLinkOnceLinkage()) ||
191 // Don't inline functions marked noinline.
192 Callee->hasFnAttr(Attribute::NoInline) || NeverInline.count(Callee))
193 return llvm::InlineCost::getNever();
195 // InlineCost - This value measures how good of an inline candidate this call
196 // site is to inline. A lower inline cost make is more likely for the call to
197 // be inlined. This value may go negative.
201 // If there is only one call of the function, and it has internal linkage,
202 // make it almost guaranteed to be inlined.
204 if (Callee->hasInternalLinkage() && Callee->hasOneUse())
207 // If this function uses the coldcc calling convention, prefer not to inline
209 if (Callee->getCallingConv() == CallingConv::Cold)
212 // If the instruction after the call, or if the normal destination of the
213 // invoke is an unreachable instruction, the function is noreturn. As such,
214 // there is little point in inlining this.
215 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
216 if (isa<UnreachableInst>(II->getNormalDest()->begin()))
218 } else if (isa<UnreachableInst>(++BasicBlock::iterator(TheCall)))
221 // Get information about the callee...
222 FunctionInfo &CalleeFI = CachedFunctionInfo[Callee];
224 // If we haven't calculated this information yet, do so now.
225 if (CalleeFI.NumBlocks == 0)
226 CalleeFI.analyzeFunction(Callee);
228 // If we should never inline this, return a huge cost.
229 if (CalleeFI.NeverInline)
230 return InlineCost::getNever();
232 if (CalleeFI.usesDynamicAlloca) {
233 // Get infomation about the caller...
234 FunctionInfo &CallerFI = CachedFunctionInfo[Caller];
236 // If we haven't calculated this information yet, do so now.
237 if (CallerFI.NumBlocks == 0)
238 CallerFI.analyzeFunction(Caller);
240 // Don't inline a callee with dynamic alloca into a caller without them.
241 // Functions containing dynamic alloca's are inefficient in various ways;
242 // don't create more inefficiency.
243 if (!CallerFI.usesDynamicAlloca)
244 return InlineCost::getNever();
247 // FIXME: It would be nice to kill off CalleeFI.NeverInline. Then we
248 // could move this up and avoid computing the FunctionInfo for
249 // things we are going to just return always inline for. This
250 // requires handling setjmp somewhere else, however.
251 if (!Callee->isDeclaration() && Callee->hasFnAttr(Attribute::AlwaysInline))
252 return InlineCost::getAlways();
254 // Add to the inline quality for properties that make the call valuable to
255 // inline. This includes factors that indicate that the result of inlining
256 // the function will be optimizable. Currently this just looks at arguments
257 // passed into the function.
260 for (CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
261 I != E; ++I, ++ArgNo) {
262 // Each argument passed in has a cost at both the caller and the callee
263 // sides. This favors functions that take many arguments over functions
264 // that take few arguments.
267 // If this is a function being passed in, it is very likely that we will be
268 // able to turn an indirect function call into a direct function call.
269 if (isa<Function>(I))
272 // If an alloca is passed in, inlining this function is likely to allow
273 // significant future optimization possibilities (like scalar promotion, and
274 // scalarization), so encourage the inlining of the function.
276 else if (isa<AllocaInst>(I)) {
277 if (ArgNo < CalleeFI.ArgumentWeights.size())
278 InlineCost -= CalleeFI.ArgumentWeights[ArgNo].AllocaWeight;
280 // If this is a constant being passed into the function, use the argument
281 // weights calculated for the callee to determine how much will be folded
282 // away with this information.
283 } else if (isa<Constant>(I)) {
284 if (ArgNo < CalleeFI.ArgumentWeights.size())
285 InlineCost -= CalleeFI.ArgumentWeights[ArgNo].ConstantWeight;
289 // Now that we have considered all of the factors that make the call site more
290 // likely to be inlined, look at factors that make us not want to inline it.
292 // Don't inline into something too big, which would make it bigger.
294 InlineCost += Caller->size()/15;
296 // Look at the size of the callee. Each instruction counts as 5.
297 InlineCost += CalleeFI.NumInsts*5;
299 return llvm::InlineCost::get(InlineCost);
302 // getInlineFudgeFactor - Return a > 1.0 factor if the inliner should use a
303 // higher threshold to determine if the function call should be inlined.
304 float InlineCostAnalyzer::getInlineFudgeFactor(CallSite CS) {
305 Function *Callee = CS.getCalledFunction();
307 // Get information about the callee...
308 FunctionInfo &CalleeFI = CachedFunctionInfo[Callee];
310 // If we haven't calculated this information yet, do so now.
311 if (CalleeFI.NumBlocks == 0)
312 CalleeFI.analyzeFunction(Callee);
315 // Single BB functions are often written to be inlined.
316 if (CalleeFI.NumBlocks == 1)
319 // Be more aggressive if the function contains a good chunk (if it mades up
320 // at least 10% of the instructions) of vector instructions.
321 if (CalleeFI.NumVectorInsts > CalleeFI.NumInsts/2)
323 else if (CalleeFI.NumVectorInsts > CalleeFI.NumInsts/10)