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 // HasAllocas - Keep track of whether or not a function contains an alloca
35 // instruction that is not in the entry block of the function. Inlining
36 // this call could cause us to blow out the stack, because the stack memory
37 // would never be released.
39 // FIXME: LLVM needs a way of dealloca'ing memory, which would make this
44 // NumInsts, NumBlocks - Keep track of how large each function is, which is
45 // used to estimate the code size cost of inlining it.
46 unsigned NumInsts, NumBlocks;
48 // ArgumentWeights - Each formal argument of the function is inspected to
49 // see if it is used in any contexts where making it a constant or alloca
50 // would reduce the code size. If so, we add some value to the argument
52 std::vector<ArgInfo> ArgumentWeights;
54 FunctionInfo() : HasAllocas(false), NumInsts(0), NumBlocks(0) {}
56 /// analyzeFunction - Fill in the current structure with information gleaned
57 /// from the specified function.
58 void analyzeFunction(Function *F);
61 class SimpleInliner : public Inliner {
62 std::map<const Function*, FunctionInfo> CachedFunctionInfo;
64 int getInlineCost(CallSite CS);
66 RegisterOpt<SimpleInliner> X("inline", "Function Integration/Inlining");
69 ModulePass *llvm::createFunctionInliningPass() { return new SimpleInliner(); }
71 // CountCodeReductionForConstant - Figure out an approximation for how many
72 // instructions will be constant folded if the specified value is constant.
74 static unsigned CountCodeReductionForConstant(Value *V) {
75 unsigned Reduction = 0;
76 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI)
77 if (isa<BranchInst>(*UI))
78 Reduction += 40; // Eliminating a conditional branch is a big win
79 else if (SwitchInst *SI = dyn_cast<SwitchInst>(*UI))
80 // Eliminating a switch is a big win, proportional to the number of edges
82 Reduction += (SI->getNumSuccessors()-1) * 40;
83 else if (CallInst *CI = dyn_cast<CallInst>(*UI)) {
84 // Turning an indirect call into a direct call is a BIG win
85 Reduction += CI->getCalledValue() == V ? 500 : 0;
86 } else if (InvokeInst *II = dyn_cast<InvokeInst>(*UI)) {
87 // Turning an indirect call into a direct call is a BIG win
88 Reduction += II->getCalledValue() == V ? 500 : 0;
90 // Figure out if this instruction will be removed due to simple constant
92 Instruction &Inst = cast<Instruction>(**UI);
93 bool AllOperandsConstant = true;
94 for (unsigned i = 0, e = Inst.getNumOperands(); i != e; ++i)
95 if (!isa<Constant>(Inst.getOperand(i)) && Inst.getOperand(i) != V) {
96 AllOperandsConstant = false;
100 if (AllOperandsConstant) {
101 // We will get to remove this instruction...
104 // And any other instructions that use it which become constants
106 Reduction += CountCodeReductionForConstant(&Inst);
113 // CountCodeReductionForAlloca - Figure out an approximation of how much smaller
114 // the function will be if it is inlined into a context where an argument
115 // becomes an alloca.
117 static unsigned CountCodeReductionForAlloca(Value *V) {
118 if (!isa<PointerType>(V->getType())) return 0; // Not a pointer
119 unsigned Reduction = 0;
120 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){
121 Instruction *I = cast<Instruction>(*UI);
122 if (isa<LoadInst>(I) || isa<StoreInst>(I))
124 else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
125 // If the GEP has variable indices, we won't be able to do much with it.
126 for (Instruction::op_iterator I = GEP->op_begin()+1, E = GEP->op_end();
128 if (!isa<Constant>(*I)) return 0;
129 Reduction += CountCodeReductionForAlloca(GEP)+15;
131 // If there is some other strange instruction, we're not going to be able
132 // to do much if we inline this.
140 /// analyzeFunction - Fill in the current structure with information gleaned
141 /// from the specified function.
142 void FunctionInfo::analyzeFunction(Function *F) {
143 unsigned NumInsts = 0, NumBlocks = 0;
145 // Look at the size of the callee. Each basic block counts as 20 units, and
146 // each instruction counts as 10.
147 for (Function::const_iterator BB = F->begin(), E = F->end(); BB != E; ++BB) {
148 for (BasicBlock::const_iterator II = BB->begin(), E = BB->end();
152 // If there is an alloca in the body of the function, we cannot currently
153 // inline the function without the risk of exploding the stack.
154 if (isa<AllocaInst>(II) && BB != F->begin()) {
156 this->NumBlocks = this->NumInsts = 1;
164 this->NumBlocks = NumBlocks;
165 this->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 for (Function::aiterator I = F->abegin(), E = F->aend(); I != E; ++I)
170 ArgumentWeights.push_back(ArgInfo(CountCodeReductionForConstant(I),
171 CountCodeReductionForAlloca(I)));
175 // getInlineCost - The heuristic used to determine if we should inline the
176 // function call or not.
178 int SimpleInliner::getInlineCost(CallSite CS) {
179 Instruction *TheCall = CS.getInstruction();
180 Function *Callee = CS.getCalledFunction();
181 const Function *Caller = TheCall->getParent()->getParent();
183 // Don't inline a directly recursive call.
184 if (Caller == Callee) return 2000000000;
186 // InlineCost - This value measures how good of an inline candidate this call
187 // site is to inline. A lower inline cost make is more likely for the call to
188 // be inlined. This value may go negative.
192 // If there is only one call of the function, and it has internal linkage,
193 // make it almost guaranteed to be inlined.
195 if (Callee->hasInternalLinkage() && Callee->hasOneUse())
198 // Get information about the callee...
199 FunctionInfo &CalleeFI = CachedFunctionInfo[Callee];
201 // If we haven't calculated this information yet, do so now.
202 if (CalleeFI.NumBlocks == 0)
203 CalleeFI.analyzeFunction(Callee);
205 // Don't inline calls to functions with allocas that are not in the entry
206 // block of the function.
207 if (CalleeFI.HasAllocas)
210 // Add to the inline quality for properties that make the call valuable to
211 // inline. This includes factors that indicate that the result of inlining
212 // the function will be optimizable. Currently this just looks at arguments
213 // passed into the function.
216 for (CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
217 I != E; ++I, ++ArgNo) {
218 // Each argument passed in has a cost at both the caller and the callee
219 // sides. This favors functions that take many arguments over functions
220 // that take few arguments.
223 // If this is a function being passed in, it is very likely that we will be
224 // able to turn an indirect function call into a direct function call.
225 if (isa<Function>(I))
228 // If an alloca is passed in, inlining this function is likely to allow
229 // significant future optimization possibilities (like scalar promotion, and
230 // scalarization), so encourage the inlining of the function.
232 else if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) {
233 if (ArgNo < CalleeFI.ArgumentWeights.size())
234 InlineCost -= CalleeFI.ArgumentWeights[ArgNo].AllocaWeight;
236 // If this is a constant being passed into the function, use the argument
237 // weights calculated for the callee to determine how much will be folded
238 // away with this information.
239 } else if (isa<Constant>(I)) {
240 if (ArgNo < CalleeFI.ArgumentWeights.size())
241 InlineCost -= CalleeFI.ArgumentWeights[ArgNo].ConstantWeight;
245 // Now that we have considered all of the factors that make the call site more
246 // likely to be inlined, look at factors that make us not want to inline it.
248 // Don't inline into something too big, which would make it bigger. Here, we
249 // count each basic block as a single unit.
251 InlineCost += Caller->size()/20;
254 // Look at the size of the callee. Each basic block counts as 20 units, and
255 // each instruction counts as 5.
256 InlineCost += CalleeFI.NumInsts*5 + CalleeFI.NumBlocks*20;