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/CallingConv.h"
16 #include "llvm/Instructions.h"
17 #include "llvm/IntrinsicInst.h"
18 #include "llvm/Function.h"
19 #include "llvm/Type.h"
20 #include "llvm/Support/CallSite.h"
21 #include "llvm/Support/Compiler.h"
22 #include "llvm/Transforms/IPO.h"
26 struct VISIBILITY_HIDDEN ArgInfo {
27 unsigned ConstantWeight;
28 unsigned AllocaWeight;
30 ArgInfo(unsigned CWeight, unsigned AWeight)
31 : ConstantWeight(CWeight), AllocaWeight(AWeight) {}
34 // FunctionInfo - For each function, calculate the size of it in blocks and
36 struct VISIBILITY_HIDDEN FunctionInfo {
37 // NumInsts, NumBlocks - Keep track of how large each function is, which is
38 // used to estimate the code size cost of inlining it.
39 unsigned NumInsts, NumBlocks;
41 // ArgumentWeights - Each formal argument of the function is inspected to
42 // see if it is used in any contexts where making it a constant or alloca
43 // would reduce the code size. If so, we add some value to the argument
45 std::vector<ArgInfo> ArgumentWeights;
47 FunctionInfo() : NumInsts(0), NumBlocks(0) {}
49 /// analyzeFunction - Fill in the current structure with information gleaned
50 /// from the specified function.
51 void analyzeFunction(Function *F);
54 class VISIBILITY_HIDDEN SimpleInliner : public Inliner {
55 std::map<const Function*, FunctionInfo> CachedFunctionInfo;
57 int getInlineCost(CallSite CS);
59 RegisterPass<SimpleInliner> X("inline", "Function Integration/Inlining");
62 Pass *llvm::createFunctionInliningPass() { return new SimpleInliner(); }
64 // CountCodeReductionForConstant - Figure out an approximation for how many
65 // instructions will be constant folded if the specified value is constant.
67 static unsigned CountCodeReductionForConstant(Value *V) {
68 unsigned Reduction = 0;
69 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI)
70 if (isa<BranchInst>(*UI))
71 Reduction += 40; // Eliminating a conditional branch is a big win
72 else if (SwitchInst *SI = dyn_cast<SwitchInst>(*UI))
73 // Eliminating a switch is a big win, proportional to the number of edges
75 Reduction += (SI->getNumSuccessors()-1) * 40;
76 else if (CallInst *CI = dyn_cast<CallInst>(*UI)) {
77 // Turning an indirect call into a direct call is a BIG win
78 Reduction += CI->getCalledValue() == V ? 500 : 0;
79 } else if (InvokeInst *II = dyn_cast<InvokeInst>(*UI)) {
80 // Turning an indirect call into a direct call is a BIG win
81 Reduction += II->getCalledValue() == V ? 500 : 0;
83 // Figure out if this instruction will be removed due to simple constant
85 Instruction &Inst = cast<Instruction>(**UI);
86 bool AllOperandsConstant = true;
87 for (unsigned i = 0, e = Inst.getNumOperands(); i != e; ++i)
88 if (!isa<Constant>(Inst.getOperand(i)) && Inst.getOperand(i) != V) {
89 AllOperandsConstant = false;
93 if (AllOperandsConstant) {
94 // We will get to remove this instruction...
97 // And any other instructions that use it which become constants
99 Reduction += CountCodeReductionForConstant(&Inst);
106 // CountCodeReductionForAlloca - Figure out an approximation of how much smaller
107 // the function will be if it is inlined into a context where an argument
108 // becomes an alloca.
110 static unsigned CountCodeReductionForAlloca(Value *V) {
111 if (!isa<PointerType>(V->getType())) return 0; // Not a pointer
112 unsigned Reduction = 0;
113 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){
114 Instruction *I = cast<Instruction>(*UI);
115 if (isa<LoadInst>(I) || isa<StoreInst>(I))
117 else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
118 // If the GEP has variable indices, we won't be able to do much with it.
119 for (Instruction::op_iterator I = GEP->op_begin()+1, E = GEP->op_end();
121 if (!isa<Constant>(*I)) return 0;
122 Reduction += CountCodeReductionForAlloca(GEP)+15;
124 // If there is some other strange instruction, we're not going to be able
125 // to do much if we inline this.
133 /// analyzeFunction - Fill in the current structure with information gleaned
134 /// from the specified function.
135 void FunctionInfo::analyzeFunction(Function *F) {
136 unsigned NumInsts = 0, NumBlocks = 0;
138 // Look at the size of the callee. Each basic block counts as 20 units, and
139 // each instruction counts as 10.
140 for (Function::const_iterator BB = F->begin(), E = F->end(); BB != E; ++BB) {
141 for (BasicBlock::const_iterator II = BB->begin(), E = BB->end();
143 if (isa<DbgInfoIntrinsic>(II)) continue; // Debug intrinsics don't count.
145 // Noop casts, including ptr <-> int, don't count.
146 if (const CastInst *CI = dyn_cast<CastInst>(II)) {
147 if (CI->isLosslessCast() || isa<IntToPtrInst>(CI) ||
148 isa<PtrToIntInst>(CI))
150 } else if (const GetElementPtrInst *GEPI =
151 dyn_cast<GetElementPtrInst>(II)) {
152 // If a GEP has all constant indices, it will probably be folded with
154 bool AllConstant = true;
155 for (unsigned i = 1, e = GEPI->getNumOperands(); i != e; ++i)
156 if (!isa<ConstantInt>(GEPI->getOperand(i))) {
160 if (AllConstant) continue;
169 this->NumBlocks = NumBlocks;
170 this->NumInsts = NumInsts;
172 // Check out all of the arguments to the function, figuring out how much
173 // code can be eliminated if one of the arguments is a constant.
174 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E; ++I)
175 ArgumentWeights.push_back(ArgInfo(CountCodeReductionForConstant(I),
176 CountCodeReductionForAlloca(I)));
180 // getInlineCost - The heuristic used to determine if we should inline the
181 // function call or not.
183 int SimpleInliner::getInlineCost(CallSite CS) {
184 Instruction *TheCall = CS.getInstruction();
185 Function *Callee = CS.getCalledFunction();
186 const Function *Caller = TheCall->getParent()->getParent();
188 // Don't inline a directly recursive call.
189 if (Caller == Callee) return 2000000000;
191 // InlineCost - This value measures how good of an inline candidate this call
192 // site is to inline. A lower inline cost make is more likely for the call to
193 // be inlined. This value may go negative.
197 // If there is only one call of the function, and it has internal linkage,
198 // make it almost guaranteed to be inlined.
200 if (Callee->hasInternalLinkage() && Callee->hasOneUse())
203 // If this function uses the coldcc calling convention, prefer not to inline
205 if (Callee->getCallingConv() == CallingConv::Cold)
208 // If the instruction after the call, or if the normal destination of the
209 // invoke is an unreachable instruction, the function is noreturn. As such,
210 // there is little point in inlining this.
211 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
212 if (isa<UnreachableInst>(II->getNormalDest()->begin()))
214 } else if (isa<UnreachableInst>(++BasicBlock::iterator(TheCall)))
217 // Get information about the callee...
218 FunctionInfo &CalleeFI = CachedFunctionInfo[Callee];
220 // If we haven't calculated this information yet, do so now.
221 if (CalleeFI.NumBlocks == 0)
222 CalleeFI.analyzeFunction(Callee);
224 // Add to the inline quality for properties that make the call valuable to
225 // inline. This includes factors that indicate that the result of inlining
226 // the function will be optimizable. Currently this just looks at arguments
227 // passed into the function.
230 for (CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
231 I != E; ++I, ++ArgNo) {
232 // Each argument passed in has a cost at both the caller and the callee
233 // sides. This favors functions that take many arguments over functions
234 // that take few arguments.
237 // If this is a function being passed in, it is very likely that we will be
238 // able to turn an indirect function call into a direct function call.
239 if (isa<Function>(I))
242 // If an alloca is passed in, inlining this function is likely to allow
243 // significant future optimization possibilities (like scalar promotion, and
244 // scalarization), so encourage the inlining of the function.
246 else if (isa<AllocaInst>(I)) {
247 if (ArgNo < CalleeFI.ArgumentWeights.size())
248 InlineCost -= CalleeFI.ArgumentWeights[ArgNo].AllocaWeight;
250 // If this is a constant being passed into the function, use the argument
251 // weights calculated for the callee to determine how much will be folded
252 // away with this information.
253 } else if (isa<Constant>(I)) {
254 if (ArgNo < CalleeFI.ArgumentWeights.size())
255 InlineCost -= CalleeFI.ArgumentWeights[ArgNo].ConstantWeight;
259 // Now that we have considered all of the factors that make the call site more
260 // likely to be inlined, look at factors that make us not want to inline it.
262 // Don't inline into something too big, which would make it bigger. Here, we
263 // count each basic block as a single unit.
265 InlineCost += Caller->size()/20;
268 // Look at the size of the callee. Each basic block counts as 20 units, and
269 // each instruction counts as 5.
270 InlineCost += CalleeFI.NumInsts*5 + CalleeFI.NumBlocks*20;