X-Git-Url: http://plrg.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FTransforms%2FIPO%2FInlineSimple.cpp;h=45609f891ed87368f65f56c2b5509cefece65195;hb=0651a407f6a408caa02b76a1cdaf0d8fa54f29fb;hp=83cfe901a09f383fa041aa304f0cd530f0e6b819;hpb=6f7426ec2e46bb19cc9f9e75f1c355b35cf12d7d;p=oota-llvm.git diff --git a/lib/Transforms/IPO/InlineSimple.cpp b/lib/Transforms/IPO/InlineSimple.cpp index 83cfe901a09..45609f891ed 100644 --- a/lib/Transforms/IPO/InlineSimple.cpp +++ b/lib/Transforms/IPO/InlineSimple.cpp @@ -2,8 +2,8 @@ // // The LLVM Compiler Infrastructure // -// This file was developed by the LLVM research group and is distributed under -// the University of Illinois Open Source License. See LICENSE.TXT for details. +// This file is distributed under the University of Illinois Open Source +// License. See LICENSE.TXT for details. // //===----------------------------------------------------------------------===// // @@ -11,309 +11,100 @@ // //===----------------------------------------------------------------------===// -#define DEBUG_TYPE "inline" -#include "llvm/CallingConv.h" -#include "llvm/Instructions.h" -#include "llvm/IntrinsicInst.h" -#include "llvm/Module.h" -#include "llvm/Type.h" +#include "llvm/Analysis/AssumptionCache.h" #include "llvm/Analysis/CallGraph.h" -#include "llvm/Support/CallSite.h" -#include "llvm/Support/Compiler.h" +#include "llvm/Analysis/InlineCost.h" +#include "llvm/Analysis/TargetLibraryInfo.h" +#include "llvm/Analysis/TargetTransformInfo.h" +#include "llvm/IR/CallSite.h" +#include "llvm/IR/CallingConv.h" +#include "llvm/IR/DataLayout.h" +#include "llvm/IR/Instructions.h" +#include "llvm/IR/IntrinsicInst.h" +#include "llvm/IR/Module.h" +#include "llvm/IR/Type.h" #include "llvm/Transforms/IPO.h" #include "llvm/Transforms/IPO/InlinerPass.h" -#include using namespace llvm; -namespace { - struct VISIBILITY_HIDDEN ArgInfo { - unsigned ConstantWeight; - unsigned AllocaWeight; - - ArgInfo(unsigned CWeight, unsigned AWeight) - : ConstantWeight(CWeight), AllocaWeight(AWeight) {} - }; - - // FunctionInfo - For each function, calculate the size of it in blocks and - // instructions. - struct VISIBILITY_HIDDEN FunctionInfo { - // NumInsts, NumBlocks - Keep track of how large each function is, which is - // used to estimate the code size cost of inlining it. - unsigned NumInsts, NumBlocks; - - // ArgumentWeights - Each formal argument of the function is inspected to - // see if it is used in any contexts where making it a constant or alloca - // would reduce the code size. If so, we add some value to the argument - // entry here. - std::vector ArgumentWeights; - - FunctionInfo() : NumInsts(0), NumBlocks(0) {} - - /// analyzeFunction - Fill in the current structure with information gleaned - /// from the specified function. - void analyzeFunction(Function *F); - }; - - class VISIBILITY_HIDDEN SimpleInliner : public Inliner { - std::map CachedFunctionInfo; - std::set NeverInline; // Functions that are never inlined - public: - SimpleInliner() : Inliner(&ID) {} - static char ID; // Pass identification, replacement for typeid - int getInlineCost(CallSite CS); - virtual bool doInitialization(CallGraph &CG); - }; - char SimpleInliner::ID = 0; - RegisterPass X("inline", "Function Integration/Inlining"); -} - -Pass *llvm::createFunctionInliningPass() { return new SimpleInliner(); } - -// CountCodeReductionForConstant - Figure out an approximation for how many -// instructions will be constant folded if the specified value is constant. -// -static unsigned CountCodeReductionForConstant(Value *V) { - unsigned Reduction = 0; - for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI) - if (isa(*UI)) - Reduction += 40; // Eliminating a conditional branch is a big win - else if (SwitchInst *SI = dyn_cast(*UI)) - // Eliminating a switch is a big win, proportional to the number of edges - // deleted. - Reduction += (SI->getNumSuccessors()-1) * 40; - else if (CallInst *CI = dyn_cast(*UI)) { - // Turning an indirect call into a direct call is a BIG win - Reduction += CI->getCalledValue() == V ? 500 : 0; - } else if (InvokeInst *II = dyn_cast(*UI)) { - // Turning an indirect call into a direct call is a BIG win - Reduction += II->getCalledValue() == V ? 500 : 0; - } else { - // Figure out if this instruction will be removed due to simple constant - // propagation. - Instruction &Inst = cast(**UI); - bool AllOperandsConstant = true; - for (unsigned i = 0, e = Inst.getNumOperands(); i != e; ++i) - if (!isa(Inst.getOperand(i)) && Inst.getOperand(i) != V) { - AllOperandsConstant = false; - break; - } - - if (AllOperandsConstant) { - // We will get to remove this instruction... - Reduction += 7; +#define DEBUG_TYPE "inline" - // And any other instructions that use it which become constants - // themselves. - Reduction += CountCodeReductionForConstant(&Inst); - } - } +namespace { - return Reduction; -} +/// \brief Actual inliner pass implementation. +/// +/// The common implementation of the inlining logic is shared between this +/// inliner pass and the always inliner pass. The two passes use different cost +/// analyses to determine when to inline. +class SimpleInliner : public Inliner { -// CountCodeReductionForAlloca - Figure out an approximation of how much smaller -// the function will be if it is inlined into a context where an argument -// becomes an alloca. -// -static unsigned CountCodeReductionForAlloca(Value *V) { - if (!isa(V->getType())) return 0; // Not a pointer - unsigned Reduction = 0; - for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){ - Instruction *I = cast(*UI); - if (isa(I) || isa(I)) - Reduction += 10; - else if (GetElementPtrInst *GEP = dyn_cast(I)) { - // If the GEP has variable indices, we won't be able to do much with it. - for (Instruction::op_iterator I = GEP->op_begin()+1, E = GEP->op_end(); - I != E; ++I) - if (!isa(*I)) return 0; - Reduction += CountCodeReductionForAlloca(GEP)+15; - } else { - // If there is some other strange instruction, we're not going to be able - // to do much if we inline this. - return 0; - } +public: + SimpleInliner() : Inliner(ID) { + initializeSimpleInlinerPass(*PassRegistry::getPassRegistry()); } - return Reduction; -} - -/// analyzeFunction - Fill in the current structure with information gleaned -/// from the specified function. -void FunctionInfo::analyzeFunction(Function *F) { - unsigned NumInsts = 0, NumBlocks = 0; + SimpleInliner(int Threshold) + : Inliner(ID, Threshold, /*InsertLifetime*/ true) { + initializeSimpleInlinerPass(*PassRegistry::getPassRegistry()); + } - // Look at the size of the callee. Each basic block counts as 20 units, and - // each instruction counts as 10. - for (Function::const_iterator BB = F->begin(), E = F->end(); BB != E; ++BB) { - for (BasicBlock::const_iterator II = BB->begin(), E = BB->end(); - II != E; ++II) { - if (isa(II)) continue; // Debug intrinsics don't count. - - // Noop casts, including ptr <-> int, don't count. - if (const CastInst *CI = dyn_cast(II)) { - if (CI->isLosslessCast() || isa(CI) || - isa(CI)) - continue; - } else if (const GetElementPtrInst *GEPI = - dyn_cast(II)) { - // If a GEP has all constant indices, it will probably be folded with - // a load/store. - bool AllConstant = true; - for (unsigned i = 1, e = GEPI->getNumOperands(); i != e; ++i) - if (!isa(GEPI->getOperand(i))) { - AllConstant = false; - break; - } - if (AllConstant) continue; - } - - ++NumInsts; - } + static char ID; // Pass identification, replacement for typeid - ++NumBlocks; + InlineCost getInlineCost(CallSite CS) override { + Function *Callee = CS.getCalledFunction(); + TargetTransformInfo &TTI = TTIWP->getTTI(*Callee); + return llvm::getInlineCost(CS, getInlineThreshold(CS), TTI, ACT); } - this->NumBlocks = NumBlocks; - this->NumInsts = NumInsts; - - // Check out all of the arguments to the function, figuring out how much - // code can be eliminated if one of the arguments is a constant. - for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E; ++I) - ArgumentWeights.push_back(ArgInfo(CountCodeReductionForConstant(I), - CountCodeReductionForAlloca(I))); + bool runOnSCC(CallGraphSCC &SCC) override; + void getAnalysisUsage(AnalysisUsage &AU) const override; + +private: + TargetTransformInfoWrapperPass *TTIWP; +}; + +static int computeThresholdFromOptLevels(unsigned OptLevel, + unsigned SizeOptLevel) { + if (OptLevel > 2) + return 275; + if (SizeOptLevel == 1) // -Os + return 75; + if (SizeOptLevel == 2) // -Oz + return 25; + return 225; } +} // end anonymous namespace -// getInlineCost - The heuristic used to determine if we should inline the -// function call or not. -// -int SimpleInliner::getInlineCost(CallSite CS) { - Instruction *TheCall = CS.getInstruction(); - Function *Callee = CS.getCalledFunction(); - const Function *Caller = TheCall->getParent()->getParent(); - - // Don't inline a directly recursive call. - if (Caller == Callee) return 2000000000; - - // Don't inline functions marked noinline - if (NeverInline.count(Callee)) return 2000000000; - - // InlineCost - This value measures how good of an inline candidate this call - // site is to inline. A lower inline cost make is more likely for the call to - // be inlined. This value may go negative. - // - int InlineCost = 0; - - // If there is only one call of the function, and it has internal linkage, - // make it almost guaranteed to be inlined. - // - if (Callee->hasInternalLinkage() && Callee->hasOneUse()) - InlineCost -= 30000; - - // If this function uses the coldcc calling convention, prefer not to inline - // it. - if (Callee->getCallingConv() == CallingConv::Cold) - InlineCost += 2000; - - // If the instruction after the call, or if the normal destination of the - // invoke is an unreachable instruction, the function is noreturn. As such, - // there is little point in inlining this. - if (InvokeInst *II = dyn_cast(TheCall)) { - if (isa(II->getNormalDest()->begin())) - InlineCost += 10000; - } else if (isa(++BasicBlock::iterator(TheCall))) - InlineCost += 10000; - - // Get information about the callee... - FunctionInfo &CalleeFI = CachedFunctionInfo[Callee]; - - // If we haven't calculated this information yet, do so now. - if (CalleeFI.NumBlocks == 0) - CalleeFI.analyzeFunction(Callee); - - // Add to the inline quality for properties that make the call valuable to - // inline. This includes factors that indicate that the result of inlining - // the function will be optimizable. Currently this just looks at arguments - // passed into the function. - // - unsigned ArgNo = 0; - for (CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end(); - I != E; ++I, ++ArgNo) { - // Each argument passed in has a cost at both the caller and the callee - // sides. This favors functions that take many arguments over functions - // that take few arguments. - InlineCost -= 20; - - // If this is a function being passed in, it is very likely that we will be - // able to turn an indirect function call into a direct function call. - if (isa(I)) - InlineCost -= 100; - - // If an alloca is passed in, inlining this function is likely to allow - // significant future optimization possibilities (like scalar promotion, and - // scalarization), so encourage the inlining of the function. - // - else if (isa(I)) { - if (ArgNo < CalleeFI.ArgumentWeights.size()) - InlineCost -= CalleeFI.ArgumentWeights[ArgNo].AllocaWeight; - - // If this is a constant being passed into the function, use the argument - // weights calculated for the callee to determine how much will be folded - // away with this information. - } else if (isa(I)) { - if (ArgNo < CalleeFI.ArgumentWeights.size()) - InlineCost -= CalleeFI.ArgumentWeights[ArgNo].ConstantWeight; - } - } - - // Now that we have considered all of the factors that make the call site more - // likely to be inlined, look at factors that make us not want to inline it. - - // Don't inline into something too big, which would make it bigger. Here, we - // count each basic block as a single unit. - // - InlineCost += Caller->size()/20; +char SimpleInliner::ID = 0; +INITIALIZE_PASS_BEGIN(SimpleInliner, "inline", + "Function Integration/Inlining", false, false) +INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker) +INITIALIZE_PASS_DEPENDENCY(CallGraphWrapperPass) +INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass) +INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass) +INITIALIZE_PASS_END(SimpleInliner, "inline", + "Function Integration/Inlining", false, false) +Pass *llvm::createFunctionInliningPass() { return new SimpleInliner(); } - // Look at the size of the callee. Each basic block counts as 20 units, and - // each instruction counts as 5. - InlineCost += CalleeFI.NumInsts*5 + CalleeFI.NumBlocks*20; - return InlineCost; +Pass *llvm::createFunctionInliningPass(int Threshold) { + return new SimpleInliner(Threshold); } -// doInitialization - Initializes the vector of functions that have been -// annotated with the noinline attribute. -bool SimpleInliner::doInitialization(CallGraph &CG) { - - Module &M = CG.getModule(); - - // Get llvm.noinline - GlobalVariable *GV = M.getNamedGlobal("llvm.noinline"); - - if (GV == 0) - return false; +Pass *llvm::createFunctionInliningPass(unsigned OptLevel, + unsigned SizeOptLevel) { + return new SimpleInliner( + computeThresholdFromOptLevels(OptLevel, SizeOptLevel)); +} - const ConstantArray *InitList = dyn_cast(GV->getInitializer()); - - if (InitList == 0) - return false; +bool SimpleInliner::runOnSCC(CallGraphSCC &SCC) { + TTIWP = &getAnalysis(); + return Inliner::runOnSCC(SCC); +} - // Iterate over each element and add to the NeverInline set - for (unsigned i = 0, e = InitList->getNumOperands(); i != e; ++i) { - - // Get Source - const Constant *Elt = InitList->getOperand(i); - - if (const ConstantExpr *CE = dyn_cast(Elt)) - if (CE->getOpcode() == Instruction::BitCast) - Elt = CE->getOperand(0); - - // Insert into set of functions to never inline - if (const Function *F = dyn_cast(Elt)) - NeverInline.insert(F); - } - - return false; +void SimpleInliner::getAnalysisUsage(AnalysisUsage &AU) const { + AU.addRequired(); + Inliner::getAnalysisUsage(AU); }