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/Support/Compiler.h"
22 #include "llvm/Transforms/IPO.h"
26 struct VISIBILITY_HIDDEN ArgInfo {
27 unsigned ConstantWeight;
28 unsigned AllocaWeight;
30 ArgInfo(unsigned CWeight, unsigned AWeight)
31 : ConstantWeight(CWeight), AllocaWeight(AWeight) {}
34 // FunctionInfo - For each function, calculate the size of it in blocks and
36 struct VISIBILITY_HIDDEN FunctionInfo {
37 // NumInsts, NumBlocks - Keep track of how large each function is, which is
38 // used to estimate the code size cost of inlining it.
39 unsigned NumInsts, NumBlocks;
41 // ArgumentWeights - Each formal argument of the function is inspected to
42 // see if it is used in any contexts where making it a constant or alloca
43 // would reduce the code size. If so, we add some value to the argument
45 std::vector<ArgInfo> ArgumentWeights;
47 FunctionInfo() : NumInsts(0), NumBlocks(0) {}
49 /// analyzeFunction - Fill in the current structure with information gleaned
50 /// from the specified function.
51 void analyzeFunction(Function *F);
54 class VISIBILITY_HIDDEN SimpleInliner : public Inliner {
55 std::map<const Function*, FunctionInfo> CachedFunctionInfo;
57 SimpleInliner() : Inliner(&ID) {}
58 static char ID; // Pass identification, replacement for typeid
59 int getInlineCost(CallSite CS);
61 char SimpleInliner::ID = 0;
62 RegisterPass<SimpleInliner> X("inline", "Function Integration/Inlining");
65 Pass *llvm::createFunctionInliningPass() { return new SimpleInliner(); }
67 // CountCodeReductionForConstant - Figure out an approximation for how many
68 // instructions will be constant folded if the specified value is constant.
70 static unsigned CountCodeReductionForConstant(Value *V) {
71 unsigned Reduction = 0;
72 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI)
73 if (isa<BranchInst>(*UI))
74 Reduction += 40; // Eliminating a conditional branch is a big win
75 else if (SwitchInst *SI = dyn_cast<SwitchInst>(*UI))
76 // Eliminating a switch is a big win, proportional to the number of edges
78 Reduction += (SI->getNumSuccessors()-1) * 40;
79 else if (CallInst *CI = dyn_cast<CallInst>(*UI)) {
80 // Turning an indirect call into a direct call is a BIG win
81 Reduction += CI->getCalledValue() == V ? 500 : 0;
82 } else if (InvokeInst *II = dyn_cast<InvokeInst>(*UI)) {
83 // Turning an indirect call into a direct call is a BIG win
84 Reduction += II->getCalledValue() == V ? 500 : 0;
86 // Figure out if this instruction will be removed due to simple constant
88 Instruction &Inst = cast<Instruction>(**UI);
89 bool AllOperandsConstant = true;
90 for (unsigned i = 0, e = Inst.getNumOperands(); i != e; ++i)
91 if (!isa<Constant>(Inst.getOperand(i)) && Inst.getOperand(i) != V) {
92 AllOperandsConstant = false;
96 if (AllOperandsConstant) {
97 // We will get to remove this instruction...
100 // And any other instructions that use it which become constants
102 Reduction += CountCodeReductionForConstant(&Inst);
109 // CountCodeReductionForAlloca - Figure out an approximation of how much smaller
110 // the function will be if it is inlined into a context where an argument
111 // becomes an alloca.
113 static unsigned CountCodeReductionForAlloca(Value *V) {
114 if (!isa<PointerType>(V->getType())) return 0; // Not a pointer
115 unsigned Reduction = 0;
116 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){
117 Instruction *I = cast<Instruction>(*UI);
118 if (isa<LoadInst>(I) || isa<StoreInst>(I))
120 else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
121 // If the GEP has variable indices, we won't be able to do much with it.
122 for (Instruction::op_iterator I = GEP->op_begin()+1, E = GEP->op_end();
124 if (!isa<Constant>(*I)) return 0;
125 Reduction += CountCodeReductionForAlloca(GEP)+15;
127 // If there is some other strange instruction, we're not going to be able
128 // to do much if we inline this.
136 /// analyzeFunction - Fill in the current structure with information gleaned
137 /// from the specified function.
138 void FunctionInfo::analyzeFunction(Function *F) {
139 unsigned NumInsts = 0, NumBlocks = 0;
141 // Look at the size of the callee. Each basic block counts as 20 units, and
142 // each instruction counts as 10.
143 for (Function::const_iterator BB = F->begin(), E = F->end(); BB != E; ++BB) {
144 for (BasicBlock::const_iterator II = BB->begin(), E = BB->end();
146 if (isa<DbgInfoIntrinsic>(II)) continue; // Debug intrinsics don't count.
148 // Noop casts, including ptr <-> int, don't count.
149 if (const CastInst *CI = dyn_cast<CastInst>(II)) {
150 if (CI->isLosslessCast() || isa<IntToPtrInst>(CI) ||
151 isa<PtrToIntInst>(CI))
153 } else if (const GetElementPtrInst *GEPI =
154 dyn_cast<GetElementPtrInst>(II)) {
155 // If a GEP has all constant indices, it will probably be folded with
157 bool AllConstant = true;
158 for (unsigned i = 1, e = GEPI->getNumOperands(); i != e; ++i)
159 if (!isa<ConstantInt>(GEPI->getOperand(i))) {
163 if (AllConstant) continue;
172 this->NumBlocks = NumBlocks;
173 this->NumInsts = NumInsts;
175 // Check out all of the arguments to the function, figuring out how much
176 // code can be eliminated if one of the arguments is a constant.
177 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E; ++I)
178 ArgumentWeights.push_back(ArgInfo(CountCodeReductionForConstant(I),
179 CountCodeReductionForAlloca(I)));
183 // getInlineCost - The heuristic used to determine if we should inline the
184 // function call or not.
186 int SimpleInliner::getInlineCost(CallSite CS) {
187 Instruction *TheCall = CS.getInstruction();
188 Function *Callee = CS.getCalledFunction();
189 const Function *Caller = TheCall->getParent()->getParent();
191 // Don't inline a directly recursive call.
192 if (Caller == Callee) return 2000000000;
194 // InlineCost - This value measures how good of an inline candidate this call
195 // site is to inline. A lower inline cost make is more likely for the call to
196 // be inlined. This value may go negative.
200 // If there is only one call of the function, and it has internal linkage,
201 // make it almost guaranteed to be inlined.
203 if (Callee->hasInternalLinkage() && Callee->hasOneUse())
206 // If this function uses the coldcc calling convention, prefer not to inline
208 if (Callee->getCallingConv() == CallingConv::Cold)
211 // If the instruction after the call, or if the normal destination of the
212 // invoke is an unreachable instruction, the function is noreturn. As such,
213 // there is little point in inlining this.
214 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
215 if (isa<UnreachableInst>(II->getNormalDest()->begin()))
217 } else if (isa<UnreachableInst>(++BasicBlock::iterator(TheCall)))
220 // Get information about the callee...
221 FunctionInfo &CalleeFI = CachedFunctionInfo[Callee];
223 // If we haven't calculated this information yet, do so now.
224 if (CalleeFI.NumBlocks == 0)
225 CalleeFI.analyzeFunction(Callee);
227 // Add to the inline quality for properties that make the call valuable to
228 // inline. This includes factors that indicate that the result of inlining
229 // the function will be optimizable. Currently this just looks at arguments
230 // passed into the function.
233 for (CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
234 I != E; ++I, ++ArgNo) {
235 // Each argument passed in has a cost at both the caller and the callee
236 // sides. This favors functions that take many arguments over functions
237 // that take few arguments.
240 // If this is a function being passed in, it is very likely that we will be
241 // able to turn an indirect function call into a direct function call.
242 if (isa<Function>(I))
245 // If an alloca is passed in, inlining this function is likely to allow
246 // significant future optimization possibilities (like scalar promotion, and
247 // scalarization), so encourage the inlining of the function.
249 else if (isa<AllocaInst>(I)) {
250 if (ArgNo < CalleeFI.ArgumentWeights.size())
251 InlineCost -= CalleeFI.ArgumentWeights[ArgNo].AllocaWeight;
253 // If this is a constant being passed into the function, use the argument
254 // weights calculated for the callee to determine how much will be folded
255 // away with this information.
256 } else if (isa<Constant>(I)) {
257 if (ArgNo < CalleeFI.ArgumentWeights.size())
258 InlineCost -= CalleeFI.ArgumentWeights[ArgNo].ConstantWeight;
262 // Now that we have considered all of the factors that make the call site more
263 // likely to be inlined, look at factors that make us not want to inline it.
265 // Don't inline into something too big, which would make it bigger. Here, we
266 // count each basic block as a single unit.
268 InlineCost += Caller->size()/20;
271 // Look at the size of the callee. Each basic block counts as 20 units, and
272 // each instruction counts as 5.
273 InlineCost += CalleeFI.NumInsts*5 + CalleeFI.NumBlocks*20;