-//===- FunctionInlining.cpp - Code to perform function inlining -----------===//
+//===- InlineSimple.cpp - Code to perform simple function inlining --------===//
//
-// This file implements inlining of functions.
+// The LLVM Compiler Infrastructure
//
-// Specifically, this:
-// * Exports functionality to inline any function call
-// * Inlines functions that consist of a single basic block
-// * Is able to inline ANY function call
-// . Has a smart heuristic for when to inline a function
+// 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.
//
-// FIXME: This pass should transform alloca instructions in the called function
-// into malloc/free pairs! Or perhaps it should refuse to inline them!
+//===----------------------------------------------------------------------===//
+//
+// This file implements bottom-up inlining of functions into callees.
//
//===----------------------------------------------------------------------===//
-#include "llvm/Transforms/IPO.h"
-#include "llvm/Transforms/Utils/Cloning.h"
-#include "llvm/Module.h"
-#include "llvm/Pass.h"
-#include "llvm/iTerminators.h"
-#include "llvm/iPHINode.h"
-#include "llvm/iOther.h"
+#include "Inliner.h"
+#include "llvm/CallingConv.h"
+#include "llvm/Instructions.h"
+#include "llvm/IntrinsicInst.h"
+#include "llvm/Function.h"
#include "llvm/Type.h"
-#include "Support/Statistic.h"
-#include <algorithm>
-
-static Statistic<> NumInlined("inline", "Number of functions inlined");
-using std::cerr;
+#include "llvm/Support/CallSite.h"
+#include "llvm/Transforms/IPO.h"
+using namespace llvm;
-// InlineFunction - This function forcibly inlines the called function 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.
-//
-// 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
-// function by one level.
-//
-bool InlineFunction(CallInst *CI) {
- assert(isa<CallInst>(CI) && "InlineFunction only works on CallInst nodes");
- assert(CI->getParent() && "Instruction not embedded in basic block!");
- assert(CI->getParent()->getParent() && "Instruction not in function!");
+namespace {
+ struct ArgInfo {
+ unsigned ConstantWeight;
+ unsigned AllocaWeight;
- const Function *CalledFunc = CI->getCalledFunction();
- if (CalledFunc == 0 || // Can't inline external function or indirect call!
- CalledFunc->isExternal()) return false;
+ ArgInfo(unsigned CWeight, unsigned AWeight)
+ : ConstantWeight(CWeight), AllocaWeight(AWeight) {}
+ };
- //cerr << "Inlining " << CalledFunc->getName() << " into "
- // << CurrentMeth->getName() << "\n";
+ // FunctionInfo - For each function, calculate the size of it in blocks and
+ // instructions.
+ struct 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;
- BasicBlock *OrigBB = CI->getParent();
+ // 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;
- // 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(CI);
- NewBB->setName("InlinedFunctionReturnNode");
+ FunctionInfo() : NumInsts(0), NumBlocks(0) {}
- // Remove (unlink) the CallInst from the start of the new basic block.
- NewBB->getInstList().remove(CI);
+ /// analyzeFunction - Fill in the current structure with information gleaned
+ /// from the specified function.
+ void analyzeFunction(Function *F);
+ };
- // 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
- // function.
- //
- PHINode *PHI = 0;
- if (!CI->use_empty()) {
- // The PHI node should go at the front of the new basic block to merge all
- // possible incoming values.
- //
- PHI = new PHINode(CalledFunc->getReturnType(), CI->getName(),
- NewBB->begin());
+ class SimpleInliner : public Inliner {
+ std::map<const Function*, FunctionInfo> CachedFunctionInfo;
+ public:
+ int getInlineCost(CallSite CS);
+ };
+ RegisterPass<SimpleInliner> X("inline", "Function Integration/Inlining");
+}
- // Anything that used the result of the function call should now use the PHI
- // node as their operand.
- //
- CI->replaceAllUsesWith(PHI);
- }
+ModulePass *llvm::createFunctionInliningPass() { return new SimpleInliner(); }
- // Get a pointer to the last basic block in the function, which will have the
- // new function inlined after it.
- //
- Function::iterator LastBlock = &OrigBB->getParent()->back();
-
- // Calculate the vector of arguments to pass into the function cloner...
- std::map<const Value*, Value*> ValueMap;
- assert((unsigned)std::distance(CalledFunc->abegin(), CalledFunc->aend()) ==
- CI->getNumOperands()-1 && "No varargs calls can be inlined yet!");
-
- unsigned i = 1;
- for (Function::const_aiterator I = CalledFunc->abegin(), E=CalledFunc->aend();
- I != E; ++I, ++i)
- ValueMap[I] = CI->getOperand(i);
-
- // Since we are now done with the CallInst, we can delete it.
- delete CI;
-
- // Make a vector to capture the return instructions in the cloned function...
- std::vector<ReturnInst*> Returns;
-
- // Populate the value map with all of the globals in the program.
- Module &M = *OrigBB->getParent()->getParent();
- for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I)
- ValueMap[I] = I;
- for (Module::giterator I = M.gbegin(), E = M.gend(); I != E; ++I)
- ValueMap[I] = I;
-
- // Do all of the hard part of cloning the callee into the caller...
- CloneFunctionInto(OrigBB->getParent(), CalledFunc, ValueMap, Returns, ".i");
-
- // Loop over all of the return instructions, turning them into unconditional
- // branches to the merge point now...
- for (unsigned i = 0, e = Returns.size(); i != e; ++i) {
- ReturnInst *RI = Returns[i];
- BasicBlock *BB = RI->getParent();
-
- // Add a branch to the merge point where the PHI node would live...
- new BranchInst(NewBB, RI);
-
- 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 function!");
- PHI->addIncoming(RI->getReturnValue(), BB);
+// 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<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);
+ }
}
- // Delete the return instruction now
- BB->getInstList().erase(RI);
- }
+ return Reduction;
+}
- // Check to see if the PHI node only has one argument. This is a common
- // case resulting from there only being a single return instruction in the
- // function call. Because this is so common, eliminate the PHI node.
- //
- if (PHI && PHI->getNumIncomingValues() == 1) {
- PHI->replaceAllUsesWith(PHI->getIncomingValue(0));
- PHI->getParent()->getInstList().erase(PHI);
+// 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<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;
+ }
}
- // Change the branch that used to go to NewBB to branch to the first basic
- // block of the inlined function.
- //
- TerminatorInst *Br = OrigBB->getTerminator();
- assert(Br && Br->getOpcode() == Instruction::Br &&
- "splitBasicBlock broken!");
- Br->setOperand(0, ++LastBlock);
- return true;
+ return Reduction;
}
-static inline bool ShouldInlineFunction(const CallInst *CI, const Function *F) {
- assert(CI->getParent() && CI->getParent()->getParent() &&
- "Call not embedded into a function!");
-
- // Don't inline a recursive call.
- if (CI->getParent()->getParent() == F) return false;
+/// 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;
+ }
- // Don't inline something too big. This is a really crappy heuristic
- if (F->size() > 3) return false;
+ ++NumBlocks;
+ }
- // Don't inline into something too big. This is a **really** crappy heuristic
- if (CI->getParent()->getParent()->size() > 10) return false;
+ this->NumBlocks = NumBlocks;
+ this->NumInsts = NumInsts;
- // Go ahead and try just about anything else.
- return true;
+ // 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)));
}
-static inline bool DoFunctionInlining(BasicBlock *BB) {
- for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
- if (CallInst *CI = dyn_cast<CallInst>(I)) {
- // Check to see if we should inline this function
- Function *F = CI->getCalledFunction();
- if (F && ShouldInlineFunction(CI, F)) {
- return InlineFunction(CI);
- }
+// 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;
+
+ // 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<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.
+ //
+ 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.
+ //
+ 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;
}
}
- return false;
-}
-// doFunctionInlining - Use a heuristic based approach to inline functions that
-// seem to look good.
-//
-static bool doFunctionInlining(Function &F) {
- bool Changed = false;
-
- // Loop through now and inline instructions a basic block at a time...
- for (Function::iterator I = F.begin(); I != F.end(); )
- if (DoFunctionInlining(I)) {
- ++NumInlined;
- Changed = true;
- } else {
- ++I;
- }
+ // 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.
- return Changed;
-}
+ // 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;
-namespace {
- struct FunctionInlining : public FunctionPass {
- virtual bool runOnFunction(Function &F) {
- return doFunctionInlining(F);
- }
- };
- RegisterOpt<FunctionInlining> X("inline", "Function Integration/Inlining");
+
+ // 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 *createFunctionInliningPass() { return new FunctionInlining(); }