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)) && Inst.getOperand(i) != V) {
82 AllOperandsConstant = false;
86 if (AllOperandsConstant) {
87 // We will get to remove this instruction...
90 // And any other instructions that use it which become constants
92 Reduction += CountCodeReductionForConstant(&Inst);
99 // CountCodeReductionForAlloca - Figure out an approximation of how much smaller
100 // the function will be if it is inlined into a context where an argument
101 // becomes an alloca.
103 static unsigned CountCodeReductionForAlloca(Value *V) {
104 if (!isa<PointerType>(V->getType())) return 0; // Not a pointer
105 unsigned Reduction = 0;
106 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){
107 Instruction *I = cast<Instruction>(*UI);
108 if (isa<LoadInst>(I) || isa<StoreInst>(I))
110 else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
111 // If the GEP has variable indices, we won't be able to do much with it.
112 for (Instruction::op_iterator I = GEP->op_begin()+1, E = GEP->op_end();
114 if (!isa<Constant>(*I)) return 0;
115 Reduction += CountCodeReductionForAlloca(GEP)+15;
117 // If there is some other strange instruction, we're not going to be able
118 // to do much if we inline this.
126 // getInlineCost - The heuristic used to determine if we should inline the
127 // function call or not.
129 int SimpleInliner::getInlineCost(CallSite CS) {
130 Instruction *TheCall = CS.getInstruction();
131 Function *Callee = CS.getCalledFunction();
132 const Function *Caller = TheCall->getParent()->getParent();
134 // Don't inline a directly recursive call.
135 if (Caller == Callee) return 2000000000;
137 // InlineCost - This value measures how good of an inline candidate this call
138 // site is to inline. A lower inline cost make is more likely for the call to
139 // be inlined. This value may go negative.
143 // If there is only one call of the function, and it has internal linkage,
144 // make it almost guaranteed to be inlined.
146 if (Callee->hasInternalLinkage() && Callee->hasOneUse())
149 // Get information about the callee...
150 FunctionInfo &CalleeFI = CachedFunctionInfo[Callee];
152 // If we haven't calculated this information yet...
153 if (CalleeFI.NumBlocks == 0) {
154 unsigned NumInsts = 0, NumBlocks = 0;
156 // Look at the size of the callee. Each basic block counts as 20 units, and
157 // each instruction counts as 10.
158 for (Function::const_iterator BB = Callee->begin(), E = Callee->end();
160 NumInsts += BB->size();
164 CalleeFI.NumBlocks = NumBlocks;
165 CalleeFI.NumInsts = NumInsts;
167 // Check out all of the arguments to the function, figuring out how much
168 // code can be eliminated if one of the arguments is a constant.
169 std::vector<ArgInfo> &ArgWeights = CalleeFI.ArgumentWeights;
171 for (Function::aiterator I = Callee->abegin(), E = Callee->aend();
173 ArgWeights.push_back(ArgInfo(CountCodeReductionForConstant(I),
174 CountCodeReductionForAlloca(I)));
178 // Add to the inline quality for properties that make the call valuable to
179 // inline. This includes factors that indicate that the result of inlining
180 // the function will be optimizable. Currently this just looks at arguments
181 // passed into the function.
184 for (CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
185 I != E; ++I, ++ArgNo) {
186 // Each argument passed in has a cost at both the caller and the callee
187 // sides. This favors functions that take many arguments over functions
188 // that take few arguments.
191 // If this is a function being passed in, it is very likely that we will be
192 // able to turn an indirect function call into a direct function call.
193 if (isa<Function>(I))
196 // If an alloca is passed in, inlining this function is likely to allow
197 // significant future optimization possibilities (like scalar promotion, and
198 // scalarization), so encourage the inlining of the function.
200 else if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) {
201 if (ArgNo < CalleeFI.ArgumentWeights.size())
202 InlineCost -= CalleeFI.ArgumentWeights[ArgNo].AllocaWeight;
204 // If this is a constant being passed into the function, use the argument
205 // weights calculated for the callee to determine how much will be folded
206 // away with this information.
207 } else if (isa<Constant>(I)) {
208 if (ArgNo < CalleeFI.ArgumentWeights.size())
209 InlineCost -= CalleeFI.ArgumentWeights[ArgNo].ConstantWeight;
213 // Now that we have considered all of the factors that make the call site more
214 // likely to be inlined, look at factors that make us not want to inline it.
216 // Don't inline into something too big, which would make it bigger. Here, we
217 // count each basic block as a single unit.
219 InlineCost += Caller->size()/20;
222 // Look at the size of the callee. Each basic block counts as 20 units, and
223 // each instruction counts as 5.
224 InlineCost += CalleeFI.NumInsts*5 + CalleeFI.NumBlocks*20;