-//===- FunctionInlining.cpp - Code to perform function inlining -----------===//
+//===- InlineSimple.cpp - Code to perform simple function inlining --------===//
+//
+// 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 implements bottom-up inlining of functions into callees.
//
//===----------------------------------------------------------------------===//
+#include "Inliner.h"
+#include "llvm/Instructions.h"
+#include "llvm/IntrinsicInst.h"
+#include "llvm/Function.h"
+#include "llvm/Type.h"
+#include "llvm/Support/CallSite.h"
#include "llvm/Transforms/IPO.h"
-#include "llvm/Transforms/Utils/Cloning.h"
-#include "llvm/Module.h"
-#include "llvm/Pass.h"
-#include "llvm/iOther.h"
-#include "llvm/iMemory.h"
-#include "Support/Statistic.h"
-#include <set>
+using namespace llvm;
namespace {
- Statistic<> NumInlined("inline", "Number of functions inlined");
-
- struct FunctionInlining : public Pass {
- virtual bool run(Module &M) {
- bool Changed = false;
- for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I)
- Changed |= doInlining(I);
- ProcessedFunctions.clear();
- return Changed;
- }
+ struct ArgInfo {
+ unsigned ConstantWeight;
+ unsigned AllocaWeight;
- private:
- std::set<Function*> ProcessedFunctions; // Prevent infinite recursion
- bool doInlining(Function *F);
+ ArgInfo(unsigned CWeight, unsigned AWeight)
+ : ConstantWeight(CWeight), AllocaWeight(AWeight) {}
};
- RegisterOpt<FunctionInlining> X("inline", "Function Integration/Inlining");
+
+ // FunctionInfo - For each function, calculate the size of it in blocks and
+ // instructions.
+ struct FunctionInfo {
+ // HasAllocas - Keep track of whether or not a function contains an alloca
+ // instruction that is not in the entry block of the function. Inlining
+ // this call could cause us to blow out the stack, because the stack memory
+ // would never be released.
+ //
+ // FIXME: LLVM needs a way of dealloca'ing memory, which would make this
+ // irrelevant!
+ //
+ bool HasAllocas;
+
+ // 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() : HasAllocas(false), NumInsts(0), NumBlocks(0) {}
+
+ /// analyzeFunction - Fill in the current structure with information gleaned
+ /// from the specified function.
+ void analyzeFunction(Function *F);
+ };
+
+ class SimpleInliner : public Inliner {
+ std::map<const Function*, FunctionInfo> CachedFunctionInfo;
+ public:
+ int getInlineCost(CallSite CS);
+ };
+ RegisterOpt<SimpleInliner> X("inline", "Function Integration/Inlining");
+}
+
+ModulePass *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<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);
+ }
+ }
+
+ return Reduction;
}
-Pass *createFunctionInliningPass() { return new FunctionInlining(); }
+// 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;
+ }
+ }
+
+ 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;
+
+ // 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)) ++NumInsts;
+
+ // If there is an alloca in the body of the function, we cannot currently
+ // inline the function without the risk of exploding the stack.
+ if (isa<AllocaInst>(II) && BB != F->begin()) {
+ HasAllocas = true;
+ this->NumBlocks = this->NumInsts = 1;
+ return;
+ }
+ }
+
+ ++NumBlocks;
+ }
+
+ 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)));
+}
-// ShouldInlineFunction - The heuristic used to determine if we should inline
-// the function call or not.
+
+// getInlineCost - The heuristic used to determine if we should inline the
+// function call or not.
//
-static inline bool ShouldInlineFunction(const CallInst *CI) {
- assert(CI->getParent() && CI->getParent()->getParent() &&
- "Call not embedded into a function!");
-
- const Function *Callee = CI->getCalledFunction();
- if (Callee == 0 || Callee->isExternal())
- return false; // Cannot inline an indirect call... or external function.
-
- // Don't inline a recursive call.
- const Function *Caller = CI->getParent()->getParent();
- if (Caller == Callee) return false;
-
- // InlineQuality - This value measures how good of an inline candidate this
- // call site is to inline. The initial value determines how aggressive the
- // inliner is. If this value is negative after the final computation,
- // inlining is not performed.
+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 InlineQuality = 200; // FIXME: This is VERY conservative
+ 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->use_size() == 1 && Callee->hasInternalLinkage())
- InlineQuality += 30000;
+ if (Callee->hasInternalLinkage() && Callee->hasOneUse())
+ InlineCost -= 30000;
+
+ // Get information about the callee...
+ FunctionInfo &CalleeFI = CachedFunctionInfo[Callee];
- // Add to the inline quality for properties that make the call valueable to
+ // If we haven't calculated this information yet, do so now.
+ if (CalleeFI.NumBlocks == 0)
+ CalleeFI.analyzeFunction(Callee);
+
+ // Don't inline calls to functions with allocas that are not in the entry
+ // block of the function.
+ if (CalleeFI.HasAllocas)
+ return 2000000000;
+
+ // 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 (User::const_op_iterator I = CI->op_begin()+1, E = CI->op_end();
- I != E; ++I){
+ 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.
- InlineQuality += 20;
+ 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))
- InlineQuality += 100;
+ 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.
+ // 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<Constant>(I) || isa<GlobalVariable>(I) || isa<AllocaInst>(I))
- InlineQuality += 60;
+ else if (AllocaInst *AI = dyn_cast<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;
+ }
}
// 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.
- // As soon as the inline quality gets negative, bail out.
-
- // 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) {
- InlineQuality -= BB->size()*10 + 20;
- if (InlineQuality < 0) return false;
- }
// Don't inline into something too big, which would make it bigger. Here, we
// count each basic block as a single unit.
- for (Function::const_iterator BB = Caller->begin(), E = Caller->end();
- BB != E; ++BB) {
- --InlineQuality;
- if (InlineQuality < 0) return false;
- }
-
- // If we get here, this call site is high enough "quality" to inline.
- DEBUG(std::cerr << "Inlining in '" << Caller->getName()
- << "', quality = " << InlineQuality << ": " << *CI);
- return true;
-}
+ //
+ InlineCost += Caller->size()/20;
-// doInlining - Use a heuristic based approach to inline functions that seem to
-// look good.
-//
-bool FunctionInlining::doInlining(Function *F) {
- // If we have already processed this function (ie, it is recursive) don't
- // revisit.
- std::set<Function*>::iterator PFI = ProcessedFunctions.lower_bound(F);
- if (PFI != ProcessedFunctions.end() && *PFI == F) return false;
-
- // Insert the function in the set so it doesn't get revisited.
- ProcessedFunctions.insert(PFI, F);
-
- bool Changed = false;
- for (Function::iterator BB = F->begin(); BB != F->end(); ++BB)
- for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ) {
- bool ShouldInc = true;
- // Found a call instruction? FIXME: This should also handle INVOKEs
- if (CallInst *CI = dyn_cast<CallInst>(I)) {
- if (Function *Callee = CI->getCalledFunction())
- doInlining(Callee); // Inline in callees before callers!
-
- // Decide whether we should inline this function...
- if (ShouldInlineFunction(CI)) {
- // Save an iterator to the instruction before the call if it exists,
- // otherwise get an iterator at the end of the block... because the
- // call will be destroyed.
- //
- BasicBlock::iterator SI;
- if (I != BB->begin()) {
- SI = I; --SI; // Instruction before the call...
- } else {
- SI = BB->end();
- }
-
- // Attempt to inline the function...
- if (InlineFunction(CI)) {
- ++NumInlined;
- Changed = true;
- // Move to instruction before the call...
- I = (SI == BB->end()) ? BB->begin() : SI;
- ShouldInc = false; // Don't increment iterator until next time
- }
- }
- }
- if (ShouldInc) ++I;
- }
-
- return Changed;
+ // 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;
}