1 //===- InlineSimple.cpp - Code to perform simple function inlining --------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by the LLVM research group and is distributed under
6 // the University of Illinois Open Source License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements bottom-up inlining of functions into callees.
12 //===----------------------------------------------------------------------===//
15 #include "llvm/CallingConv.h"
16 #include "llvm/Instructions.h"
17 #include "llvm/IntrinsicInst.h"
18 #include "llvm/Function.h"
19 #include "llvm/Type.h"
20 #include "llvm/Support/CallSite.h"
21 #include "llvm/Transforms/IPO.h"
26 unsigned ConstantWeight;
27 unsigned AllocaWeight;
29 ArgInfo(unsigned CWeight, unsigned AWeight)
30 : ConstantWeight(CWeight), AllocaWeight(AWeight) {}
33 // FunctionInfo - For each function, calculate the size of it in blocks and
36 // NumInsts, NumBlocks - Keep track of how large each function is, which is
37 // used to estimate the code size cost of inlining it.
38 unsigned NumInsts, NumBlocks;
40 // ArgumentWeights - Each formal argument of the function is inspected to
41 // see if it is used in any contexts where making it a constant or alloca
42 // would reduce the code size. If so, we add some value to the argument
44 std::vector<ArgInfo> ArgumentWeights;
46 FunctionInfo() : NumInsts(0), NumBlocks(0) {}
48 /// analyzeFunction - Fill in the current structure with information gleaned
49 /// from the specified function.
50 void analyzeFunction(Function *F);
53 class SimpleInliner : public Inliner {
54 std::map<const Function*, FunctionInfo> CachedFunctionInfo;
56 int getInlineCost(CallSite CS);
58 RegisterPass<SimpleInliner> X("inline", "Function Integration/Inlining");
61 ModulePass *llvm::createFunctionInliningPass() { return new SimpleInliner(); }
63 // CountCodeReductionForConstant - Figure out an approximation for how many
64 // instructions will be constant folded if the specified value is constant.
66 static unsigned CountCodeReductionForConstant(Value *V) {
67 unsigned Reduction = 0;
68 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI)
69 if (isa<BranchInst>(*UI))
70 Reduction += 40; // Eliminating a conditional branch is a big win
71 else if (SwitchInst *SI = dyn_cast<SwitchInst>(*UI))
72 // Eliminating a switch is a big win, proportional to the number of edges
74 Reduction += (SI->getNumSuccessors()-1) * 40;
75 else if (CallInst *CI = dyn_cast<CallInst>(*UI)) {
76 // Turning an indirect call into a direct call is a BIG win
77 Reduction += CI->getCalledValue() == V ? 500 : 0;
78 } else if (InvokeInst *II = dyn_cast<InvokeInst>(*UI)) {
79 // Turning an indirect call into a direct call is a BIG win
80 Reduction += II->getCalledValue() == V ? 500 : 0;
82 // Figure out if this instruction will be removed due to simple constant
84 Instruction &Inst = cast<Instruction>(**UI);
85 bool AllOperandsConstant = true;
86 for (unsigned i = 0, e = Inst.getNumOperands(); i != e; ++i)
87 if (!isa<Constant>(Inst.getOperand(i)) && Inst.getOperand(i) != V) {
88 AllOperandsConstant = false;
92 if (AllOperandsConstant) {
93 // We will get to remove this instruction...
96 // And any other instructions that use it which become constants
98 Reduction += CountCodeReductionForConstant(&Inst);
105 // CountCodeReductionForAlloca - Figure out an approximation of how much smaller
106 // the function will be if it is inlined into a context where an argument
107 // becomes an alloca.
109 static unsigned CountCodeReductionForAlloca(Value *V) {
110 if (!isa<PointerType>(V->getType())) return 0; // Not a pointer
111 unsigned Reduction = 0;
112 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){
113 Instruction *I = cast<Instruction>(*UI);
114 if (isa<LoadInst>(I) || isa<StoreInst>(I))
116 else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
117 // If the GEP has variable indices, we won't be able to do much with it.
118 for (Instruction::op_iterator I = GEP->op_begin()+1, E = GEP->op_end();
120 if (!isa<Constant>(*I)) return 0;
121 Reduction += CountCodeReductionForAlloca(GEP)+15;
123 // If there is some other strange instruction, we're not going to be able
124 // to do much if we inline this.
132 /// analyzeFunction - Fill in the current structure with information gleaned
133 /// from the specified function.
134 void FunctionInfo::analyzeFunction(Function *F) {
135 unsigned NumInsts = 0, NumBlocks = 0;
137 // Look at the size of the callee. Each basic block counts as 20 units, and
138 // each instruction counts as 10.
139 for (Function::const_iterator BB = F->begin(), E = F->end(); BB != E; ++BB) {
140 for (BasicBlock::const_iterator II = BB->begin(), E = BB->end();
142 if (isa<DbgInfoIntrinsic>(II)) continue; // Debug intrinsics don't count.
144 // Noop casts don't count.
145 if (const CastInst *CI = dyn_cast<CastInst>(II)) {
146 const Type *OpTy = CI->getOperand(0)->getType();
147 if (CI->getType()->isLosslesslyConvertibleTo(OpTy))
149 if ((isa<PointerType>(CI->getType()) && OpTy->isInteger()) ||
150 (isa<PointerType>(OpTy) && CI->getType()->isInteger()))
151 continue; // ptr <-> int is *probably* noop cast.
152 } else if (const GetElementPtrInst *GEPI =
153 dyn_cast<GetElementPtrInst>(II)) {
154 // If a GEP has all constant indices, it will probably be folded with
156 bool AllConstant = true;
157 for (unsigned i = 1, e = GEPI->getNumOperands(); i != e; ++i)
158 if (!isa<ConstantInt>(GEPI->getOperand(i))) {
162 if (AllConstant) continue;
171 this->NumBlocks = NumBlocks;
172 this->NumInsts = NumInsts;
174 // Check out all of the arguments to the function, figuring out how much
175 // code can be eliminated if one of the arguments is a constant.
176 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E; ++I)
177 ArgumentWeights.push_back(ArgInfo(CountCodeReductionForConstant(I),
178 CountCodeReductionForAlloca(I)));
182 // getInlineCost - The heuristic used to determine if we should inline the
183 // function call or not.
185 int SimpleInliner::getInlineCost(CallSite CS) {
186 Instruction *TheCall = CS.getInstruction();
187 Function *Callee = CS.getCalledFunction();
188 const Function *Caller = TheCall->getParent()->getParent();
190 // Don't inline a directly recursive call.
191 if (Caller == Callee) return 2000000000;
193 // InlineCost - This value measures how good of an inline candidate this call
194 // site is to inline. A lower inline cost make is more likely for the call to
195 // be inlined. This value may go negative.
199 // If there is only one call of the function, and it has internal linkage,
200 // make it almost guaranteed to be inlined.
202 if (Callee->hasInternalLinkage() && Callee->hasOneUse())
205 // If this function uses the coldcc calling convention, prefer not to inline
207 if (Callee->getCallingConv() == CallingConv::Cold)
210 // If the instruction after the call, or if the normal destination of the
211 // invoke is an unreachable instruction, the function is noreturn. As such,
212 // there is little point in inlining this.
213 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
214 if (isa<UnreachableInst>(II->getNormalDest()->begin()))
216 } else if (isa<UnreachableInst>(++BasicBlock::iterator(TheCall)))
219 // Get information about the callee...
220 FunctionInfo &CalleeFI = CachedFunctionInfo[Callee];
222 // If we haven't calculated this information yet, do so now.
223 if (CalleeFI.NumBlocks == 0)
224 CalleeFI.analyzeFunction(Callee);
226 // Add to the inline quality for properties that make the call valuable to
227 // inline. This includes factors that indicate that the result of inlining
228 // the function will be optimizable. Currently this just looks at arguments
229 // passed into the function.
232 for (CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
233 I != E; ++I, ++ArgNo) {
234 // Each argument passed in has a cost at both the caller and the callee
235 // sides. This favors functions that take many arguments over functions
236 // that take few arguments.
239 // If this is a function being passed in, it is very likely that we will be
240 // able to turn an indirect function call into a direct function call.
241 if (isa<Function>(I))
244 // If an alloca is passed in, inlining this function is likely to allow
245 // significant future optimization possibilities (like scalar promotion, and
246 // scalarization), so encourage the inlining of the function.
248 else if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) {
249 if (ArgNo < CalleeFI.ArgumentWeights.size())
250 InlineCost -= CalleeFI.ArgumentWeights[ArgNo].AllocaWeight;
252 // If this is a constant being passed into the function, use the argument
253 // weights calculated for the callee to determine how much will be folded
254 // away with this information.
255 } else if (isa<Constant>(I)) {
256 if (ArgNo < CalleeFI.ArgumentWeights.size())
257 InlineCost -= CalleeFI.ArgumentWeights[ArgNo].ConstantWeight;
261 // Now that we have considered all of the factors that make the call site more
262 // likely to be inlined, look at factors that make us not want to inline it.
264 // Don't inline into something too big, which would make it bigger. Here, we
265 // count each basic block as a single unit.
267 InlineCost += Caller->size()/20;
270 // Look at the size of the callee. Each basic block counts as 20 units, and
271 // each instruction counts as 5.
272 InlineCost += CalleeFI.NumInsts*5 + CalleeFI.NumBlocks*20;