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/IntrinsicInst.h"
17 #include "llvm/Function.h"
18 #include "llvm/Type.h"
19 #include "llvm/Support/CallSite.h"
20 #include "llvm/Transforms/IPO.h"
25 unsigned ConstantWeight;
26 unsigned AllocaWeight;
28 ArgInfo(unsigned CWeight, unsigned AWeight)
29 : ConstantWeight(CWeight), AllocaWeight(AWeight) {}
32 // FunctionInfo - For each function, calculate the size of it in blocks and
35 // HasAllocas - Keep track of whether or not a function contains an alloca
36 // instruction that is not in the entry block of the function. Inlining
37 // this call could cause us to blow out the stack, because the stack memory
38 // would never be released.
40 // FIXME: LLVM needs a way of dealloca'ing memory, which would make this
45 // NumInsts, NumBlocks - Keep track of how large each function is, which is
46 // used to estimate the code size cost of inlining it.
47 unsigned NumInsts, NumBlocks;
49 // ArgumentWeights - Each formal argument of the function is inspected to
50 // see if it is used in any contexts where making it a constant or alloca
51 // would reduce the code size. If so, we add some value to the argument
53 std::vector<ArgInfo> ArgumentWeights;
55 FunctionInfo() : HasAllocas(false), NumInsts(0), NumBlocks(0) {}
57 /// analyzeFunction - Fill in the current structure with information gleaned
58 /// from the specified function.
59 void analyzeFunction(Function *F);
62 class SimpleInliner : public Inliner {
63 std::map<const Function*, FunctionInfo> CachedFunctionInfo;
65 int getInlineCost(CallSite CS);
67 RegisterOpt<SimpleInliner> X("inline", "Function Integration/Inlining");
70 ModulePass *llvm::createFunctionInliningPass() { return new SimpleInliner(); }
72 // CountCodeReductionForConstant - Figure out an approximation for how many
73 // instructions will be constant folded if the specified value is constant.
75 static unsigned CountCodeReductionForConstant(Value *V) {
76 unsigned Reduction = 0;
77 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI)
78 if (isa<BranchInst>(*UI))
79 Reduction += 40; // Eliminating a conditional branch is a big win
80 else if (SwitchInst *SI = dyn_cast<SwitchInst>(*UI))
81 // Eliminating a switch is a big win, proportional to the number of edges
83 Reduction += (SI->getNumSuccessors()-1) * 40;
84 else if (CallInst *CI = dyn_cast<CallInst>(*UI)) {
85 // Turning an indirect call into a direct call is a BIG win
86 Reduction += CI->getCalledValue() == V ? 500 : 0;
87 } else if (InvokeInst *II = dyn_cast<InvokeInst>(*UI)) {
88 // Turning an indirect call into a direct call is a BIG win
89 Reduction += II->getCalledValue() == V ? 500 : 0;
91 // Figure out if this instruction will be removed due to simple constant
93 Instruction &Inst = cast<Instruction>(**UI);
94 bool AllOperandsConstant = true;
95 for (unsigned i = 0, e = Inst.getNumOperands(); i != e; ++i)
96 if (!isa<Constant>(Inst.getOperand(i)) && Inst.getOperand(i) != V) {
97 AllOperandsConstant = false;
101 if (AllOperandsConstant) {
102 // We will get to remove this instruction...
105 // And any other instructions that use it which become constants
107 Reduction += CountCodeReductionForConstant(&Inst);
114 // CountCodeReductionForAlloca - Figure out an approximation of how much smaller
115 // the function will be if it is inlined into a context where an argument
116 // becomes an alloca.
118 static unsigned CountCodeReductionForAlloca(Value *V) {
119 if (!isa<PointerType>(V->getType())) return 0; // Not a pointer
120 unsigned Reduction = 0;
121 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){
122 Instruction *I = cast<Instruction>(*UI);
123 if (isa<LoadInst>(I) || isa<StoreInst>(I))
125 else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
126 // If the GEP has variable indices, we won't be able to do much with it.
127 for (Instruction::op_iterator I = GEP->op_begin()+1, E = GEP->op_end();
129 if (!isa<Constant>(*I)) return 0;
130 Reduction += CountCodeReductionForAlloca(GEP)+15;
132 // If there is some other strange instruction, we're not going to be able
133 // to do much if we inline this.
141 /// analyzeFunction - Fill in the current structure with information gleaned
142 /// from the specified function.
143 void FunctionInfo::analyzeFunction(Function *F) {
144 unsigned NumInsts = 0, NumBlocks = 0;
146 // Look at the size of the callee. Each basic block counts as 20 units, and
147 // each instruction counts as 10.
148 for (Function::const_iterator BB = F->begin(), E = F->end(); BB != E; ++BB) {
149 for (BasicBlock::const_iterator II = BB->begin(), E = BB->end();
151 if (!isa<DbgInfoIntrinsic>(II)) ++NumInsts;
153 // If there is an alloca in the body of the function, we cannot currently
154 // inline the function without the risk of exploding the stack.
155 if (isa<AllocaInst>(II) && BB != F->begin()) {
157 this->NumBlocks = this->NumInsts = 1;
165 this->NumBlocks = NumBlocks;
166 this->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 for (Function::aiterator I = F->abegin(), E = F->aend(); I != E; ++I)
171 ArgumentWeights.push_back(ArgInfo(CountCodeReductionForConstant(I),
172 CountCodeReductionForAlloca(I)));
176 // getInlineCost - The heuristic used to determine if we should inline the
177 // function call or not.
179 int SimpleInliner::getInlineCost(CallSite CS) {
180 Instruction *TheCall = CS.getInstruction();
181 Function *Callee = CS.getCalledFunction();
182 const Function *Caller = TheCall->getParent()->getParent();
184 // Don't inline a directly recursive call.
185 if (Caller == Callee) return 2000000000;
187 // InlineCost - This value measures how good of an inline candidate this call
188 // site is to inline. A lower inline cost make is more likely for the call to
189 // be inlined. This value may go negative.
193 // If there is only one call of the function, and it has internal linkage,
194 // make it almost guaranteed to be inlined.
196 if (Callee->hasInternalLinkage() && Callee->hasOneUse())
199 // Get information about the callee...
200 FunctionInfo &CalleeFI = CachedFunctionInfo[Callee];
202 // If we haven't calculated this information yet, do so now.
203 if (CalleeFI.NumBlocks == 0)
204 CalleeFI.analyzeFunction(Callee);
206 // Don't inline calls to functions with allocas that are not in the entry
207 // block of the function.
208 if (CalleeFI.HasAllocas)
211 // Add to the inline quality for properties that make the call valuable to
212 // inline. This includes factors that indicate that the result of inlining
213 // the function will be optimizable. Currently this just looks at arguments
214 // passed into the function.
217 for (CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
218 I != E; ++I, ++ArgNo) {
219 // Each argument passed in has a cost at both the caller and the callee
220 // sides. This favors functions that take many arguments over functions
221 // that take few arguments.
224 // If this is a function being passed in, it is very likely that we will be
225 // able to turn an indirect function call into a direct function call.
226 if (isa<Function>(I))
229 // If an alloca is passed in, inlining this function is likely to allow
230 // significant future optimization possibilities (like scalar promotion, and
231 // scalarization), so encourage the inlining of the function.
233 else if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) {
234 if (ArgNo < CalleeFI.ArgumentWeights.size())
235 InlineCost -= CalleeFI.ArgumentWeights[ArgNo].AllocaWeight;
237 // If this is a constant being passed into the function, use the argument
238 // weights calculated for the callee to determine how much will be folded
239 // away with this information.
240 } else if (isa<Constant>(I)) {
241 if (ArgNo < CalleeFI.ArgumentWeights.size())
242 InlineCost -= CalleeFI.ArgumentWeights[ArgNo].ConstantWeight;
246 // Now that we have considered all of the factors that make the call site more
247 // likely to be inlined, look at factors that make us not want to inline it.
249 // Don't inline into something too big, which would make it bigger. Here, we
250 // count each basic block as a single unit.
252 InlineCost += Caller->size()/20;
255 // Look at the size of the callee. Each basic block counts as 20 units, and
256 // each instruction counts as 5.
257 InlineCost += CalleeFI.NumInsts*5 + CalleeFI.NumBlocks*20;