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/Transforms/IPO.h"
26 unsigned ConstantWeight;
27 unsigned AllocaWeight;
29 ArgInfo(unsigned CWeight, unsigned AWeight)
30 : ConstantWeight(CWeight), AllocaWeight(AWeight) {}
33 // FunctionInfo - For each function, calculate the size of it in blocks and
36 // HasAllocas - Keep track of whether or not a function contains an alloca
37 // instruction that is not in the entry block of the function. Inlining
38 // this call could cause us to blow out the stack, because the stack memory
39 // would never be released.
41 // FIXME: LLVM needs a way of dealloca'ing memory, which would make this
46 // NumInsts, NumBlocks - Keep track of how large each function is, which is
47 // used to estimate the code size cost of inlining it.
48 unsigned NumInsts, NumBlocks;
50 // ArgumentWeights - Each formal argument of the function is inspected to
51 // see if it is used in any contexts where making it a constant or alloca
52 // would reduce the code size. If so, we add some value to the argument
54 std::vector<ArgInfo> ArgumentWeights;
56 FunctionInfo() : HasAllocas(false), NumInsts(0), NumBlocks(0) {}
58 /// analyzeFunction - Fill in the current structure with information gleaned
59 /// from the specified function.
60 void analyzeFunction(Function *F);
63 class SimpleInliner : public Inliner {
64 std::map<const Function*, FunctionInfo> CachedFunctionInfo;
66 int getInlineCost(CallSite CS);
68 RegisterOpt<SimpleInliner> X("inline", "Function Integration/Inlining");
71 ModulePass *llvm::createFunctionInliningPass() { return new SimpleInliner(); }
73 // CountCodeReductionForConstant - Figure out an approximation for how many
74 // instructions will be constant folded if the specified value is constant.
76 static unsigned CountCodeReductionForConstant(Value *V) {
77 unsigned Reduction = 0;
78 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI)
79 if (isa<BranchInst>(*UI))
80 Reduction += 40; // Eliminating a conditional branch is a big win
81 else if (SwitchInst *SI = dyn_cast<SwitchInst>(*UI))
82 // Eliminating a switch is a big win, proportional to the number of edges
84 Reduction += (SI->getNumSuccessors()-1) * 40;
85 else if (CallInst *CI = dyn_cast<CallInst>(*UI)) {
86 // Turning an indirect call into a direct call is a BIG win
87 Reduction += CI->getCalledValue() == V ? 500 : 0;
88 } else if (InvokeInst *II = dyn_cast<InvokeInst>(*UI)) {
89 // Turning an indirect call into a direct call is a BIG win
90 Reduction += II->getCalledValue() == V ? 500 : 0;
92 // Figure out if this instruction will be removed due to simple constant
94 Instruction &Inst = cast<Instruction>(**UI);
95 bool AllOperandsConstant = true;
96 for (unsigned i = 0, e = Inst.getNumOperands(); i != e; ++i)
97 if (!isa<Constant>(Inst.getOperand(i)) && Inst.getOperand(i) != V) {
98 AllOperandsConstant = false;
102 if (AllOperandsConstant) {
103 // We will get to remove this instruction...
106 // And any other instructions that use it which become constants
108 Reduction += CountCodeReductionForConstant(&Inst);
115 // CountCodeReductionForAlloca - Figure out an approximation of how much smaller
116 // the function will be if it is inlined into a context where an argument
117 // becomes an alloca.
119 static unsigned CountCodeReductionForAlloca(Value *V) {
120 if (!isa<PointerType>(V->getType())) return 0; // Not a pointer
121 unsigned Reduction = 0;
122 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){
123 Instruction *I = cast<Instruction>(*UI);
124 if (isa<LoadInst>(I) || isa<StoreInst>(I))
126 else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
127 // If the GEP has variable indices, we won't be able to do much with it.
128 for (Instruction::op_iterator I = GEP->op_begin()+1, E = GEP->op_end();
130 if (!isa<Constant>(*I)) return 0;
131 Reduction += CountCodeReductionForAlloca(GEP)+15;
133 // If there is some other strange instruction, we're not going to be able
134 // to do much if we inline this.
142 /// analyzeFunction - Fill in the current structure with information gleaned
143 /// from the specified function.
144 void FunctionInfo::analyzeFunction(Function *F) {
145 unsigned NumInsts = 0, NumBlocks = 0;
147 // Look at the size of the callee. Each basic block counts as 20 units, and
148 // each instruction counts as 10.
149 for (Function::const_iterator BB = F->begin(), E = F->end(); BB != E; ++BB) {
150 for (BasicBlock::const_iterator II = BB->begin(), E = BB->end();
152 if (!isa<DbgInfoIntrinsic>(II)) ++NumInsts;
154 // If there is an alloca in the body of the function, we cannot currently
155 // inline the function without the risk of exploding the stack.
156 if (isa<AllocaInst>(II) && BB != F->begin()) {
158 this->NumBlocks = this->NumInsts = 1;
166 this->NumBlocks = NumBlocks;
167 this->NumInsts = NumInsts;
169 // Check out all of the arguments to the function, figuring out how much
170 // code can be eliminated if one of the arguments is a constant.
171 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E; ++I)
172 ArgumentWeights.push_back(ArgInfo(CountCodeReductionForConstant(I),
173 CountCodeReductionForAlloca(I)));
177 // getInlineCost - The heuristic used to determine if we should inline the
178 // function call or not.
180 int SimpleInliner::getInlineCost(CallSite CS) {
181 Instruction *TheCall = CS.getInstruction();
182 Function *Callee = CS.getCalledFunction();
183 const Function *Caller = TheCall->getParent()->getParent();
185 // Don't inline a directly recursive call.
186 if (Caller == Callee) return 2000000000;
188 // InlineCost - This value measures how good of an inline candidate this call
189 // site is to inline. A lower inline cost make is more likely for the call to
190 // be inlined. This value may go negative.
194 // If there is only one call of the function, and it has internal linkage,
195 // make it almost guaranteed to be inlined.
197 if (Callee->hasInternalLinkage() && Callee->hasOneUse())
200 // If this function uses the coldcc calling convention, prefer not to inline
202 if (Callee->getCallingConv() == CallingConv::Cold)
205 // If the instruction after the call, or if the normal destination of the
206 // invoke is an unreachable instruction, the function is noreturn. As such,
207 // there is little point in inlining this.
208 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
209 if (isa<UnreachableInst>(II->getNormalDest()->begin()))
211 } else if (isa<UnreachableInst>(++BasicBlock::iterator(TheCall)))
214 // Get information about the callee...
215 FunctionInfo &CalleeFI = CachedFunctionInfo[Callee];
217 // If we haven't calculated this information yet, do so now.
218 if (CalleeFI.NumBlocks == 0)
219 CalleeFI.analyzeFunction(Callee);
221 // Don't inline calls to functions with allocas that are not in the entry
222 // block of the function.
223 if (CalleeFI.HasAllocas)
226 // Add to the inline quality for properties that make the call valuable to
227 // inline. This includes factors that indicate that the result of inlining
228 // the function will be optimizable. Currently this just looks at arguments
229 // passed into the function.
232 for (CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
233 I != E; ++I, ++ArgNo) {
234 // Each argument passed in has a cost at both the caller and the callee
235 // sides. This favors functions that take many arguments over functions
236 // that take few arguments.
239 // If this is a function being passed in, it is very likely that we will be
240 // able to turn an indirect function call into a direct function call.
241 if (isa<Function>(I))
244 // If an alloca is passed in, inlining this function is likely to allow
245 // significant future optimization possibilities (like scalar promotion, and
246 // scalarization), so encourage the inlining of the function.
248 else if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) {
249 if (ArgNo < CalleeFI.ArgumentWeights.size())
250 InlineCost -= CalleeFI.ArgumentWeights[ArgNo].AllocaWeight;
252 // If this is a constant being passed into the function, use the argument
253 // weights calculated for the callee to determine how much will be folded
254 // away with this information.
255 } else if (isa<Constant>(I)) {
256 if (ArgNo < CalleeFI.ArgumentWeights.size())
257 InlineCost -= CalleeFI.ArgumentWeights[ArgNo].ConstantWeight;
261 // Now that we have considered all of the factors that make the call site more
262 // likely to be inlined, look at factors that make us not want to inline it.
264 // Don't inline into something too big, which would make it bigger. Here, we
265 // count each basic block as a single unit.
267 InlineCost += Caller->size()/20;
270 // Look at the size of the callee. Each basic block counts as 20 units, and
271 // each instruction counts as 5.
272 InlineCost += CalleeFI.NumInsts*5 + CalleeFI.NumBlocks*20;