-//===- FunctionInlining.cpp - Code to perform function inlining -----------===//
+//===- InlineSimple.cpp - Code to perform simple function inlining --------===//
//
-// This file implements inlining of functions.
-//
-// 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
-//
-// Notice that:
-// * This pass 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.
-//
-// FIXME: This pass should transform alloca instructions in the called function
-// into malloc/free pairs!
+// This file implements bottom-up inlining of functions into callees.
//
//===----------------------------------------------------------------------===//
-#include "llvm/Transforms/MethodInlining.h"
-#include "llvm/Module.h"
+#include "Inliner.h"
#include "llvm/Function.h"
-#include "llvm/Pass.h"
-#include "llvm/iTerminators.h"
-#include "llvm/iPHINode.h"
-#include "llvm/iOther.h"
-#include "llvm/Type.h"
-#include "llvm/Argument.h"
-#include <algorithm>
-#include <map>
-#include <iostream>
-using std::cerr;
+#include "llvm/iMemory.h"
+#include "llvm/Support/CallSite.h"
+#include "llvm/Transforms/IPO.h"
-// RemapInstruction - Convert the instruction operands from referencing the
-// current values into those specified by ValueMap.
-//
-static inline void RemapInstruction(Instruction *I,
- std::map<const Value *, Value*> &ValueMap) {
+namespace {
+ // FunctionInfo - For each function, calculate the size of it in blocks and
+ // instructions.
+ struct FunctionInfo {
+ unsigned NumInsts, NumBlocks;
- for (unsigned op = 0, E = I->getNumOperands(); op != E; ++op) {
- const Value *Op = I->getOperand(op);
- Value *V = ValueMap[Op];
- if (!V && (isa<GlobalValue>(Op) || isa<Constant>(Op)))
- continue; // Globals and constants don't get relocated
+ FunctionInfo() : NumInsts(0), NumBlocks(0) {}
+ };
- if (!V) {
- cerr << "Val = \n" << Op << "Addr = " << (void*)Op;
- cerr << "\nInst = " << I;
- }
- assert(V && "Referenced value not in value map!");
- I->setOperand(op, V);
- }
+ class SimpleInliner : public Inliner {
+ std::map<const Function*, FunctionInfo> CachedFunctionInfo;
+ public:
+ int getInlineCost(CallSite CS);
+ };
+ RegisterOpt<SimpleInliner> X("inline", "Function Integration/Inlining");
}
-// 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(BasicBlock::iterator CIIt) {
- assert(isa<CallInst>(*CIIt) && "InlineFunction only works on CallInst nodes");
- assert((*CIIt)->getParent() && "Instruction not embedded in basic block!");
- assert((*CIIt)->getParent()->getParent() && "Instruction not in function!");
-
- CallInst *CI = cast<CallInst>(*CIIt);
- const Function *CalledMeth = CI->getCalledFunction();
- if (CalledMeth == 0 || // Can't inline external function or indirect call!
- CalledMeth->isExternal()) return false;
-
- //cerr << "Inlining " << CalledMeth->getName() << " into "
- // << CurrentMeth->getName() << "\n";
-
- BasicBlock *OrigBB = CI->getParent();
-
- // 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);
- NewBB->setName("InlinedFunctionReturnNode");
+Pass *createFunctionInliningPass() { return new SimpleInliner(); }
- // Remove (unlink) the CallInst from the start of the new basic block.
- NewBB->getInstList().remove(CI);
-
- // 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 (CalledMeth->getReturnType() != Type::VoidTy) {
- PHI = new PHINode(CalledMeth->getReturnType(), CI->getName());
-
- // The PHI node should go at the front of the new basic block to merge all
- // possible incoming values.
- //
- NewBB->getInstList().push_front(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();
+ const Function *Callee = CS.getCalledFunction();
+ const Function *Caller = TheCall->getParent()->getParent();
- // Anything that used the result of the function call should now use the PHI
- // node as their operand.
- //
- CI->replaceAllUsesWith(PHI);
- }
+ // Don't inline a directly recursive call.
+ if (Caller == Callee) return 2000000000;
- // Keep a mapping between the original function's values and the new
- // duplicated code's values. This includes all of: Function arguments,
- // instruction values, constant pool entries, and basic blocks.
+ // 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.
//
- std::map<const Value *, Value*> ValueMap;
+ int InlineCost = 0;
- // Add the function arguments to the mapping: (start counting at 1 to skip the
- // function reference itself)
+ // If there is only one call of the function, and it has internal linkage,
+ // make it almost guaranteed to be inlined.
//
- Function::ArgumentListType::const_iterator PTI =
- CalledMeth->getArgumentList().begin();
- for (unsigned a = 1, E = CI->getNumOperands(); a != E; ++a, ++PTI)
- ValueMap[*PTI] = CI->getOperand(a);
-
- ValueMap[NewBB] = NewBB; // Returns get converted to reference NewBB
+ if (Callee->hasInternalLinkage() && Callee->hasOneUse())
+ InlineCost -= 30000;
- // Loop over all of the basic blocks in the function, inlining them as
- // appropriate. Keep track of the first basic block of the function...
+ // 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 (Function::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());
- if (BB->hasName()) IBB->setName(BB->getName()+".i"); // .i = inlined once
-
- ValueMap[BB] = 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!
+ for (CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
+ I != E; ++I) {
+ // 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 a constant, global variable or alloca is passed in, inlining this
+ // function is likely to allow significant future optimization possibilities
+ // (constant propagation, scalar promotion, and scalarization), so encourage
+ // the inlining of the function.
//
- 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.
- if ((*II)->hasName())
- NewInst->setName((*II)->getName()+".i"); // .i = inlined once
- }
-
- // Copy over the terminator now...
- switch (TI->getOpcode()) {
- case Instruction::Ret: {
- const ReturnInst *RI = cast<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 function!");
- PHI->addIncoming((Value*)RI->getReturnValue(), cast<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 << "FunctionInlining: Don't know how to handle terminator: " << TI;
- abort();
- }
- }
-
-
- // Loop over all of the instructions in the function, fixing up operand
- // references as we go. This uses ValueMap to do all the hard work.
- //
- for (Function::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...
- //
- for (BasicBlock::iterator II = NBB->begin(); II != NBB->end(); II++)
- RemapInstruction(*II, ValueMap);
+ else if (isa<Constant>(I) || isa<GlobalVariable>(I) || isa<AllocaInst>(I))
+ InlineCost -= 60;
}
- if (PHI) RemapInstruction(PHI, ValueMap); // Fix the PHI node also...
-
- // 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, ValueMap[CalledMeth->front()]);
-
- // Since we are now done with the CallInst, we can finally delete it.
- delete CI;
- return true;
-}
-
-bool InlineFunction(CallInst *CI) {
- assert(CI->getParent() && "CallInst not embeded in BasicBlock!");
- BasicBlock *PBB = CI->getParent();
-
- BasicBlock::iterator CallIt = find(PBB->begin(), PBB->end(), CI);
-
- assert(CallIt != PBB->end() &&
- "CallInst has parent that doesn't contain CallInst?!?");
- return InlineFunction(CallIt);
-}
-
-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;
-
- // Don't inline something too big. This is a really crappy heuristic
- if (F->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;
-}
+ // 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.
+ FunctionInfo &CalleeFI = CachedFunctionInfo[Callee];
+ // If we haven't calculated this information yet...
+ if (CalleeFI.NumBlocks == 0) {
+ unsigned NumInsts = 0, NumBlocks = 0;
-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(I);
+ // 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 = Callee->begin(), E = Callee->end();
+ BB != E; ++BB) {
+ NumInsts += BB->size();
+ NumBlocks++;
}
+ CalleeFI.NumBlocks = NumBlocks;
+ CalleeFI.NumInsts = NumInsts;
}
- return false;
-}
-// doFunctionInlining - Use a heuristic based approach to inline functions that
-// seem to look good.
-//
-static bool doFunctionInlining(Function *F) {
- bool Changed = false;
+ // 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()*2;
- // Loop through now and inline instructions a basic block at a time...
- for (Function::iterator I = F->begin(); I != F->end(); )
- if (DoFunctionInlining(*I)) {
- Changed = true;
- // Iterator is now invalidated by new basic blocks inserted
- I = F->begin();
- } else {
- ++I;
- }
- return Changed;
+ // Look at the size of the callee. Each basic block counts as 20 units, and
+ // each instruction counts as 10.
+ InlineCost += CalleeFI.NumInsts*10 + CalleeFI.NumBlocks*20;
+ return InlineCost;
}
-
-namespace {
- struct FunctionInlining : public FunctionPass {
- virtual bool runOnFunction(Function *F) {
- return doFunctionInlining(F);
- }
- };
-}
-
-Pass *createFunctionInliningPass() { return new FunctionInlining(); }