-//===- MethodInlining.cpp - Code to perform method inlining ---------------===//
+//===- InlineSimple.cpp - Code to perform simple function inlining --------===//
//
-// This file implements inlining of methods.
+// The LLVM Compiler Infrastructure
//
-// Specifically, this:
-// * Exports functionality to inline any method call
-// * Inlines methods that consist of a single basic block
-// * Is able to inline ANY method call
-// . Has a smart heuristic for when to inline a method
+// 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.
//
-// Notice that:
-// * This pass has a habit of introducing duplicated constant pool entries,
-// and also opens up a lot of opportunities for constant propogation. It is
-// a good idea to to run a constant propogation pass, then a DCE pass
-// sometime after running this pass.
+//===----------------------------------------------------------------------===//
//
-// TODO: Currently this throws away all of the symbol names in the method being
-// inlined to try to avoid name clashes. Use a name if it's not taken
+// This file implements bottom-up inlining of functions into callees.
//
//===----------------------------------------------------------------------===//
-#include "llvm/Optimizations/MethodInlining.h"
+#define DEBUG_TYPE "inline"
+#include "llvm/CallingConv.h"
+#include "llvm/Instructions.h"
+#include "llvm/IntrinsicInst.h"
#include "llvm/Module.h"
-#include "llvm/Method.h"
-#include "llvm/iTerminators.h"
-#include "llvm/iOther.h"
-#include <algorithm>
-#include <map>
-
-#include "llvm/Assembly/Writer.h"
+#include "llvm/Type.h"
+#include "llvm/Analysis/CallGraph.h"
+#include "llvm/Support/CallSite.h"
+#include "llvm/Support/Compiler.h"
+#include "llvm/Transforms/IPO.h"
+#include "llvm/Transforms/IPO/InlinerPass.h"
+#include <set>
+
+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<ArgInfo> 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<const Function*, FunctionInfo> CachedFunctionInfo;
+ std::set<const Function*> 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<SimpleInliner> X("inline", "Function Integration/Inlining");
+}
-using namespace opt;
+Pass *llvm::createFunctionInliningPass() { return new SimpleInliner(); }
-// RemapInstruction - Convert the instruction operands from referencing the
-// current values into those specified by ValueMap.
+// CountCodeReductionForConstant - Figure out an approximation for how many
+// instructions will be constant folded if the specified value is constant.
//
-static inline void RemapInstruction(Instruction *I,
- map<const Value *, Value*> &ValueMap) {
-
- for (unsigned op = 0; const Value *Op = I->getOperand(op); ++op) {
- Value *V = ValueMap[Op];
- if (!V && Op->isMethod())
- continue; // Methods don't get relocated
-
- if (!V) {
- cerr << "Val = " << endl << Op << "Addr = " << (void*)Op << endl;
- cerr << "Inst = " << I;
+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<BranchInst>(*UI))
+ Reduction += 40; // Eliminating a conditional branch is a big win
+ else if (SwitchInst *SI = dyn_cast<SwitchInst>(*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<CallInst>(*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<InvokeInst>(*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<Instruction>(**UI);
+ bool AllOperandsConstant = true;
+ for (unsigned i = 0, e = Inst.getNumOperands(); i != e; ++i)
+ if (!isa<Constant>(Inst.getOperand(i)) && Inst.getOperand(i) != V) {
+ AllOperandsConstant = false;
+ break;
+ }
+
+ if (AllOperandsConstant) {
+ // We will get to remove this instruction...
+ Reduction += 7;
+
+ // And any other instructions that use it which become constants
+ // themselves.
+ Reduction += CountCodeReductionForConstant(&Inst);
+ }
}
- assert(V && "Referenced value not in value map!");
- I->setOperand(op, V);
- }
+
+ return Reduction;
}
-// InlineMethod - This function forcibly inlines the called method into the
-// basic block of the caller. This returns false if it is not possible to
-// inline this call. The program is still in a well defined state if this
-// occurs though.
+// 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.
//
-// Note that this only does one level of inlining. For example, if the
-// instruction 'call B' is inlined, and 'B' calls 'C', then the call to 'C' now
-// exists in the instruction stream. Similiarly this will inline a recursive
-// method by one level.
-//
-bool opt::InlineMethod(BasicBlock::iterator CIIt) {
- assert((*CIIt)->getInstType() == Instruction::Call &&
- "InlineMethod only works on CallInst nodes!");
- assert((*CIIt)->getParent() && "Instruction not embedded in basic block!");
- assert((*CIIt)->getParent()->getParent() && "Instruction not in method!");
-
- CallInst *CI = (CallInst*)*CIIt;
- const Method *CalledMeth = CI->getCalledMethod();
- Method *CurrentMeth = CI->getParent()->getParent();
+static unsigned CountCodeReductionForAlloca(Value *V) {
+ if (!isa<PointerType>(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<Instruction>(*UI);
+ if (isa<LoadInst>(I) || isa<StoreInst>(I))
+ Reduction += 10;
+ else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(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<Constant>(*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;
+ }
+ }
- //cerr << "Inlining " << CalledMeth->getName() << " into "
- // << CurrentMeth->getName() << endl;
+ return Reduction;
+}
- BasicBlock *OrigBB = CI->getParent();
+/// analyzeFunction - Fill in the current structure with information gleaned
+/// from the specified function.
+void FunctionInfo::analyzeFunction(Function *F) {
+ unsigned NumInsts = 0, NumBlocks = 0;
+
+ // 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<DbgInfoIntrinsic>(II)) continue; // Debug intrinsics don't count.
+
+ // Noop casts, including ptr <-> int, don't count.
+ if (const CastInst *CI = dyn_cast<CastInst>(II)) {
+ if (CI->isLosslessCast() || isa<IntToPtrInst>(CI) ||
+ isa<PtrToIntInst>(CI))
+ continue;
+ } else if (const GetElementPtrInst *GEPI =
+ dyn_cast<GetElementPtrInst>(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<ConstantInt>(GEPI->getOperand(i))) {
+ AllConstant = false;
+ break;
+ }
+ if (AllConstant) continue;
+ }
+
+ ++NumInsts;
+ }
- // Call splitBasicBlock - The original basic block now ends at the instruction
- // immediately before the call. The original basic block now ends with an
- // unconditional branch to NewBB, and NewBB starts with the call instruction.
- //
- BasicBlock *NewBB = OrigBB->splitBasicBlock(CIIt);
+ ++NumBlocks;
+ }
- // Remove (unlink) the CallInst from the start of the new basic block.
- NewBB->getInstList().remove(CI);
+ this->NumBlocks = NumBlocks;
+ this->NumInsts = NumInsts;
- // If we have a return value generated by this call, convert it into a PHI
- // node that gets values from each of the old RET instructions in the original
- // method.
- //
- PHINode *PHI = 0;
- if (CalledMeth->getReturnType() != Type::VoidTy) {
- PHI = new PHINode(CalledMeth->getReturnType(), CI->getName());
+ // 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)));
+}
- // The PHI node should go at the front of the new basic block to merge all
- // possible incoming values.
- //
- NewBB->getInstList().push_front(PHI);
- // Anything that used the result of the function call should now use the PHI
- // node as their operand.
- //
- CI->replaceAllUsesWith(PHI);
- }
+// 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();
- // Keep a mapping between the original method's values and the new duplicated
- // code's values. This includes all of: Method arguments, instruction values,
- // constant pool entries, and basic blocks.
- //
- map<const Value *, Value*> ValueMap;
+ // Don't inline a directly recursive call.
+ if (Caller == Callee) return 2000000000;
- // Add the method arguments to the mapping: (start counting at 1 to skip the
- // method reference itself)
- //
- Method::ArgumentListType::const_iterator PTI =
- CalledMeth->getArgumentList().begin();
- for (unsigned a = 1; Value *Operand = CI->getOperand(a); ++a, ++PTI) {
- ValueMap[*PTI] = Operand;
- }
+ // Don't inline functions marked noinline
+ if (NeverInline.count(Callee)) return 2000000000;
-
- ValueMap[NewBB] = NewBB; // Returns get converted to reference NewBB
-
- // Loop over all of the basic blocks in the method, inlining them as
- // appropriate. Keep track of the first basic block of the method...
+ // 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.
//
- for (Method::const_iterator BI = CalledMeth->begin();
- BI != CalledMeth->end(); ++BI) {
- const BasicBlock *BB = *BI;
- assert(BB->getTerminator() && "BasicBlock doesn't have terminator!?!?");
-
- // Create a new basic block to copy instructions into!
- BasicBlock *IBB = new BasicBlock("", NewBB->getParent());
-
- ValueMap[*BI] = IBB; // Add basic block mapping.
-
- // Make sure to capture the mapping that a return will use...
- // TODO: This assumes that the RET is returning a value computed in the same
- // basic block as the return was issued from!
- //
- const TerminatorInst *TI = BB->getTerminator();
-
- // Loop over all instructions copying them over...
- Instruction *NewInst;
- for (BasicBlock::const_iterator II = BB->begin();
- II != (BB->end()-1); ++II) {
- IBB->getInstList().push_back((NewInst = (*II)->clone()));
- ValueMap[*II] = NewInst; // Add instruction map to value.
- }
+ int InlineCost = 0;
- // Copy over the terminator now...
- switch (TI->getInstType()) {
- case Instruction::Ret: {
- const ReturnInst *RI = (const ReturnInst*)TI;
-
- if (PHI) { // The PHI node should include this value!
- assert(RI->getReturnValue() && "Ret should have value!");
- assert(RI->getReturnValue()->getType() == PHI->getType() &&
- "Ret value not consistent in method!");
- PHI->addIncoming((Value*)RI->getReturnValue(), (BasicBlock*)BB);
- }
-
- // Add a branch to the code that was after the original Call.
- IBB->getInstList().push_back(new BranchInst(NewBB));
- break;
- }
- case Instruction::Br:
- IBB->getInstList().push_back(TI->clone());
- break;
-
- default:
- cerr << "MethodInlining: Don't know how to handle terminator: " << TI;
- abort();
- }
- }
-
-
- // Copy over the constant pool...
+ // If there is only one call of the function, and it has internal linkage,
+ // make it almost guaranteed to be inlined.
//
- const ConstantPool &CP = CalledMeth->getConstantPool();
- ConstantPool &NewCP = CurrentMeth->getConstantPool();
- for (ConstantPool::plane_const_iterator PI = CP.begin(); PI != CP.end(); ++PI){
- ConstantPool::PlaneType &Plane = **PI;
- for (ConstantPool::PlaneType::const_iterator I = Plane.begin();
- I != Plane.end(); ++I) {
- ConstPoolVal *NewVal = (*I)->clone(); // Copy existing constant
- NewCP.insert(NewVal); // Insert the new copy into local const pool
- ValueMap[*I] = NewVal; // Keep track of constant value mappings
- }
- }
-
- // Loop over all of the instructions in the method, fixing up operand
- // references as we go. This uses ValueMap to do all the hard work.
+ 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<InvokeInst>(TheCall)) {
+ if (isa<UnreachableInst>(II->getNormalDest()->begin()))
+ InlineCost += 10000;
+ } else if (isa<UnreachableInst>(++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.
//
- for (Method::const_iterator BI = CalledMeth->begin();
- BI != CalledMeth->end(); ++BI) {
- const BasicBlock *BB = *BI;
- BasicBlock *NBB = (BasicBlock*)ValueMap[BB];
-
- // Loop over all instructions, fixing each one as we find it...
+ 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<Function>(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.
//
- for (BasicBlock::iterator II = NBB->begin(); II != NBB->end(); II++)
- RemapInstruction(*II, ValueMap);
+ else if (isa<AllocaInst>(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<Constant>(I)) {
+ if (ArgNo < CalleeFI.ArgumentWeights.size())
+ InlineCost -= CalleeFI.ArgumentWeights[ArgNo].ConstantWeight;
+ }
}
- if (PHI) RemapInstruction(PHI, ValueMap); // Fix the PHI node also...
+ // 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.
- // Change the branch that used to go to NewBB to branch to the first basic
- // block of the inlined method.
+ // Don't inline into something too big, which would make it bigger. Here, we
+ // count each basic block as a single unit.
//
- TerminatorInst *Br = OrigBB->getTerminator();
- assert(Br && Br->getInstType() == Instruction::Br &&
- "splitBasicBlock broken!");
- Br->setOperand(0, ValueMap[CalledMeth->front()]);
-
- // Since we are now done with the CallInst, we can finally delete it.
- delete CI;
- return true;
-}
-
-bool opt::InlineMethod(CallInst *CI) {
- assert(CI->getParent() && "CallInst not embeded in BasicBlock!");
- BasicBlock *PBB = CI->getParent();
+ InlineCost += Caller->size()/20;
- BasicBlock::iterator CallIt = find(PBB->begin(), PBB->end(), CI);
- assert(CallIt != PBB->end() &&
- "CallInst has parent that doesn't contain CallInst?!?");
- return InlineMethod(CallIt);
-}
-
-static inline bool ShouldInlineMethod(const CallInst *CI, const Method *M) {
- assert(CI->getParent() && CI->getParent()->getParent() &&
- "Call not embedded into a method!");
-
- // Don't inline a recursive call.
- if (CI->getParent()->getParent() == M) return false;
-
- // Don't inline something too big. This is a really crappy heuristic
- if (M->size() > 3) return false;
-
- // Don't inline into something too big. This is a **really** crappy heuristic
- if (CI->getParent()->getParent()->size() > 10) return false;
-
- // Go ahead and try just about anything else.
- return true;
+ // 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;
}
+// 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;
-static inline bool DoMethodInlining(BasicBlock *BB) {
- for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
- if ((*I)->getInstType() == Instruction::Call) {
- // Check to see if we should inline this method
- CallInst *CI = (CallInst*)*I;
- Method *M = CI->getCalledMethod();
- if (ShouldInlineMethod(CI, M))
- return InlineMethod(I);
- }
+ const ConstantArray *InitList = dyn_cast<ConstantArray>(GV->getInitializer());
+
+ if (InitList == 0)
+ return false;
+
+ // 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<ConstantExpr>(Elt))
+ if (CE->getOpcode() == Instruction::BitCast)
+ Elt = CE->getOperand(0);
+
+ // Insert into set of functions to never inline
+ if (const Function *F = dyn_cast<Function>(Elt))
+ NeverInline.insert(F);
}
+
return false;
}
-
-bool opt::DoMethodInlining(Method *M) {
- bool Changed = false;
-
- // Loop through now and inline instructions a basic block at a time...
- for (Method::iterator I = M->begin(); I != M->end(); )
- if (DoMethodInlining(*I)) {
- Changed = true;
- // Iterator is now invalidated by new basic blocks inserted
- I = M->begin();
- } else {
- ++I;
- }
-
- return Changed;
-}