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/Instructions.h"
16 #include "llvm/Function.h"
17 #include "llvm/Type.h"
18 #include "llvm/Support/CallSite.h"
19 #include "llvm/Transforms/IPO.h"
24 unsigned ConstantWeight;
25 unsigned AllocaWeight;
27 ArgInfo(unsigned CWeight, unsigned AWeight)
28 : ConstantWeight(CWeight), AllocaWeight(AWeight) {}
31 // FunctionInfo - For each function, calculate the size of it in blocks and
34 // NumInsts, NumBlocks - Keep track of how large each function is, which is
35 // used to estimate the code size cost of inlining it.
36 unsigned NumInsts, NumBlocks;
38 // ArgumentWeights - Each formal argument of the function is inspected to
39 // see if it is used in any contexts where making it a constant or alloca
40 // would reduce the code size. If so, we add some value to the argument
42 std::vector<ArgInfo> ArgumentWeights;
44 FunctionInfo() : NumInsts(0), NumBlocks(0) {}
47 class SimpleInliner : public Inliner {
48 std::map<const Function*, FunctionInfo> CachedFunctionInfo;
50 int getInlineCost(CallSite CS);
52 RegisterOpt<SimpleInliner> X("inline", "Function Integration/Inlining");
55 Pass *llvm::createFunctionInliningPass() { return new SimpleInliner(); }
57 // CountCodeReductionForConstant - Figure out an approximation for how many
58 // instructions will be constant folded if the specified value is constant.
60 static unsigned CountCodeReductionForConstant(Value *V) {
61 unsigned Reduction = 0;
62 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI)
63 if (isa<BranchInst>(*UI))
64 Reduction += 40; // Eliminating a conditional branch is a big win
65 else if (SwitchInst *SI = dyn_cast<SwitchInst>(*UI))
66 // Eliminating a switch is a big win, proportional to the number of edges
68 Reduction += (SI->getNumSuccessors()-1) * 40;
69 else if (CallInst *CI = dyn_cast<CallInst>(*UI)) {
70 // Turning an indirect call into a direct call is a BIG win
71 Reduction += CI->getCalledValue() == V ? 500 : 0;
72 } else if (InvokeInst *II = dyn_cast<InvokeInst>(*UI)) {
73 // Turning an indirect call into a direct call is a BIG win
74 Reduction += II->getCalledValue() == V ? 500 : 0;
76 // Figure out if this instruction will be removed due to simple constant
78 Instruction &Inst = cast<Instruction>(**UI);
79 bool AllOperandsConstant = true;
80 for (unsigned i = 0, e = Inst.getNumOperands(); i != e; ++i)
81 if (!isa<Constant>(Inst.getOperand(i)) &&
82 !isa<GlobalValue>(Inst.getOperand(i)) && Inst.getOperand(i) != V) {
83 AllOperandsConstant = false;
87 if (AllOperandsConstant) {
88 // We will get to remove this instruction...
91 // And any other instructions that use it which become constants
93 Reduction += CountCodeReductionForConstant(&Inst);
100 // CountCodeReductionForAlloca - Figure out an approximation of how much smaller
101 // the function will be if it is inlined into a context where an argument
102 // becomes an alloca.
104 static unsigned CountCodeReductionForAlloca(Value *V) {
105 if (!isa<PointerType>(V->getType())) return 0; // Not a pointer
106 unsigned Reduction = 0;
107 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){
108 Instruction *I = cast<Instruction>(*UI);
109 if (isa<LoadInst>(I) || isa<StoreInst>(I))
111 else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
112 // If the GEP has variable indices, we won't be able to do much with it.
113 for (Instruction::op_iterator I = GEP->op_begin()+1, E = GEP->op_end();
115 if (!isa<Constant>(*I)) return 0;
116 Reduction += CountCodeReductionForAlloca(GEP)+15;
118 // If there is some other strange instruction, we're not going to be able
119 // to do much if we inline this.
127 // getInlineCost - The heuristic used to determine if we should inline the
128 // function call or not.
130 int SimpleInliner::getInlineCost(CallSite CS) {
131 Instruction *TheCall = CS.getInstruction();
132 Function *Callee = CS.getCalledFunction();
133 const Function *Caller = TheCall->getParent()->getParent();
135 // Don't inline a directly recursive call.
136 if (Caller == Callee) return 2000000000;
138 // InlineCost - This value measures how good of an inline candidate this call
139 // site is to inline. A lower inline cost make is more likely for the call to
140 // be inlined. This value may go negative.
144 // If there is only one call of the function, and it has internal linkage,
145 // make it almost guaranteed to be inlined.
147 if (Callee->hasInternalLinkage() && Callee->hasOneUse())
150 // Get information about the callee...
151 FunctionInfo &CalleeFI = CachedFunctionInfo[Callee];
153 // If we haven't calculated this information yet...
154 if (CalleeFI.NumBlocks == 0) {
155 unsigned NumInsts = 0, NumBlocks = 0;
157 // Look at the size of the callee. Each basic block counts as 20 units, and
158 // each instruction counts as 10.
159 for (Function::const_iterator BB = Callee->begin(), E = Callee->end();
161 NumInsts += BB->size();
165 CalleeFI.NumBlocks = NumBlocks;
166 CalleeFI.NumInsts = NumInsts;
168 // Check out all of the arguments to the function, figuring out how much
169 // code can be eliminated if one of the arguments is a constant.
170 std::vector<ArgInfo> &ArgWeights = CalleeFI.ArgumentWeights;
172 for (Function::aiterator I = Callee->abegin(), E = Callee->aend();
174 ArgWeights.push_back(ArgInfo(CountCodeReductionForConstant(I),
175 CountCodeReductionForAlloca(I)));
179 // Add to the inline quality for properties that make the call valuable to
180 // inline. This includes factors that indicate that the result of inlining
181 // the function will be optimizable. Currently this just looks at arguments
182 // passed into the function.
185 for (CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
186 I != E; ++I, ++ArgNo) {
187 // Each argument passed in has a cost at both the caller and the callee
188 // sides. This favors functions that take many arguments over functions
189 // that take few arguments.
192 // If this is a function being passed in, it is very likely that we will be
193 // able to turn an indirect function call into a direct function call.
194 if (isa<Function>(I))
197 // If an alloca is passed in, inlining this function is likely to allow
198 // significant future optimization possibilities (like scalar promotion, and
199 // scalarization), so encourage the inlining of the function.
201 else if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) {
202 if (ArgNo < CalleeFI.ArgumentWeights.size())
203 InlineCost -= CalleeFI.ArgumentWeights[ArgNo].AllocaWeight;
205 // If this is a constant being passed into the function, use the argument
206 // weights calculated for the callee to determine how much will be folded
207 // away with this information.
208 } else if (isa<Constant>(I) || isa<GlobalVariable>(I)) {
209 if (ArgNo < CalleeFI.ArgumentWeights.size())
210 InlineCost -= CalleeFI.ArgumentWeights[ArgNo].ConstantWeight;
214 // Now that we have considered all of the factors that make the call site more
215 // likely to be inlined, look at factors that make us not want to inline it.
217 // Don't inline into something too big, which would make it bigger. Here, we
218 // count each basic block as a single unit.
220 InlineCost += Caller->size()/20;
223 // Look at the size of the callee. Each basic block counts as 20 units, and
224 // each instruction counts as 5.
225 InlineCost += CalleeFI.NumInsts*5 + CalleeFI.NumBlocks*20;