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/Support/CallSite.h"
18 #include "llvm/Transforms/IPO.h"
22 // FunctionInfo - For each function, calculate the size of it in blocks and
25 // NumInsts, NumBlocks - Keep track of how large each function is, which is
26 // used to estimate the code size cost of inlining it.
27 unsigned NumInsts, NumBlocks;
29 // ConstantArgumentWeights - Each formal argument of the function is
30 // inspected to see if it is used in any contexts where making it a constant
31 // would reduce the code size. If so, we add some value to the argument
33 std::vector<unsigned> ConstantArgumentWeights;
35 FunctionInfo() : NumInsts(0), NumBlocks(0) {}
38 class SimpleInliner : public Inliner {
39 std::map<const Function*, FunctionInfo> CachedFunctionInfo;
41 int getInlineCost(CallSite CS);
43 RegisterOpt<SimpleInliner> X("inline", "Function Integration/Inlining");
46 Pass *llvm::createFunctionInliningPass() { return new SimpleInliner(); }
48 // CountCodeReductionForConstant - Figure out an approximation for how many
49 // instructions will be constant folded if the specified value is constant.
51 static unsigned CountCodeReductionForConstant(Value *V) {
52 unsigned Reduction = 0;
53 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI)
54 if (isa<BranchInst>(*UI))
55 Reduction += 40; // Eliminating a conditional branch is a big win
56 else if (SwitchInst *SI = dyn_cast<SwitchInst>(*UI))
57 // Eliminating a switch is a big win, proportional to the number of edges
59 Reduction += (SI->getNumSuccessors()-1) * 40;
60 else if (CallInst *CI = dyn_cast<CallInst>(*UI)) {
61 // Turning an indirect call into a direct call is a BIG win
62 Reduction += CI->getCalledValue() == V ? 500 : 0;
63 } else if (InvokeInst *II = dyn_cast<InvokeInst>(*UI)) {
64 // Turning an indirect call into a direct call is a BIG win
65 Reduction += CI->getCalledValue() == V ? 500 : 0;
67 // Figure out if this instruction will be removed due to simple constant
69 Instruction &Inst = cast<Instruction>(**UI);
70 bool AllOperandsConstant = true;
71 for (unsigned i = 0, e = Inst.getNumOperands(); i != e; ++i)
72 if (!isa<Constant>(Inst.getOperand(i)) &&
73 !isa<GlobalValue>(Inst.getOperand(i)) && Inst.getOperand(i) != V) {
74 AllOperandsConstant = false;
78 if (AllOperandsConstant) {
79 // We will get to remove this instruction...
82 // And any other instructions that use it which become constants
84 Reduction += CountCodeReductionForConstant(&Inst);
91 // getInlineCost - The heuristic used to determine if we should inline the
92 // function call or not.
94 int SimpleInliner::getInlineCost(CallSite CS) {
95 Instruction *TheCall = CS.getInstruction();
96 Function *Callee = CS.getCalledFunction();
97 const Function *Caller = TheCall->getParent()->getParent();
99 // Don't inline a directly recursive call.
100 if (Caller == Callee) return 2000000000;
102 // InlineCost - This value measures how good of an inline candidate this call
103 // site is to inline. A lower inline cost make is more likely for the call to
104 // be inlined. This value may go negative.
108 // If there is only one call of the function, and it has internal linkage,
109 // make it almost guaranteed to be inlined.
111 if (Callee->hasInternalLinkage() && Callee->hasOneUse())
114 // Get information about the callee...
115 FunctionInfo &CalleeFI = CachedFunctionInfo[Callee];
117 // If we haven't calculated this information yet...
118 if (CalleeFI.NumBlocks == 0) {
119 unsigned NumInsts = 0, NumBlocks = 0;
121 // Look at the size of the callee. Each basic block counts as 20 units, and
122 // each instruction counts as 10.
123 for (Function::const_iterator BB = Callee->begin(), E = Callee->end();
125 NumInsts += BB->size();
129 CalleeFI.NumBlocks = NumBlocks;
130 CalleeFI.NumInsts = NumInsts;
132 // Check out all of the arguments to the function, figuring out how much
133 // code can be eliminated if one of the arguments is a constant.
134 std::vector<unsigned> &ArgWeights = CalleeFI.ConstantArgumentWeights;
136 for (Function::aiterator I = Callee->abegin(), E = Callee->aend();
138 ArgWeights.push_back(CountCodeReductionForConstant(I));
142 // Add to the inline quality for properties that make the call valuable to
143 // inline. This includes factors that indicate that the result of inlining
144 // the function will be optimizable. Currently this just looks at arguments
145 // passed into the function.
148 for (CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
149 I != E; ++I, ++ArgNo) {
150 // Each argument passed in has a cost at both the caller and the callee
151 // sides. This favors functions that take many arguments over functions
152 // that take few arguments.
155 // If this is a function being passed in, it is very likely that we will be
156 // able to turn an indirect function call into a direct function call.
157 if (isa<Function>(I))
160 // If an alloca is passed in, inlining this function is likely to allow
161 // significant future optimization possibilities (like scalar promotion, and
162 // scalarization), so encourage the inlining of the function.
164 else if (isa<AllocaInst>(I))
167 // If this is a constant being passed into the function, use the argument
168 // weights calculated for the callee to determine how much will be folded
169 // away with this information.
170 else if (isa<Constant>(I) || isa<GlobalVariable>(I)) {
171 if (ArgNo < CalleeFI.ConstantArgumentWeights.size())
172 InlineCost -= CalleeFI.ConstantArgumentWeights[ArgNo];
176 // Now that we have considered all of the factors that make the call site more
177 // likely to be inlined, look at factors that make us not want to inline it.
179 // Don't inline into something too big, which would make it bigger. Here, we
180 // count each basic block as a single unit.
181 InlineCost += Caller->size()*2;
184 // Look at the size of the callee. Each basic block counts as 20 units, and
185 // each instruction counts as 5.
186 InlineCost += CalleeFI.NumInsts*5 + CalleeFI.NumBlocks*20;