X-Git-Url: http://demsky.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FAnalysis%2FInlineCost.cpp;h=b103897977b30e0f2b1bec30be0797911d10b747;hb=c0362d5c6e0066c741bce056a65d8b4c026de19f;hp=627137159706097616ff93f8af2ac191e363f9d8;hpb=752e2590585227f6f10e80978f587915d9adb2ad;p=oota-llvm.git diff --git a/lib/Analysis/InlineCost.cpp b/lib/Analysis/InlineCost.cpp index 62713715970..b103897977b 100644 --- a/lib/Analysis/InlineCost.cpp +++ b/lib/Analysis/InlineCost.cpp @@ -18,96 +18,6 @@ #include "llvm/ADT/SmallPtrSet.h" using namespace llvm; -// CountCodeReductionForConstant - Figure out an approximation for how many -// instructions will be constant folded if the specified value is constant. -// -unsigned InlineCostAnalyzer::FunctionInfo:: -CountCodeReductionForConstant(Value *V) { - unsigned Reduction = 0; - for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){ - User *U = *UI; - if (isa(U) || isa(U)) { - // We will be able to eliminate all but one of the successors. - const TerminatorInst &TI = cast(*U); - const unsigned NumSucc = TI.getNumSuccessors(); - unsigned Instrs = 0; - for (unsigned I = 0; I != NumSucc; ++I) - Instrs += Metrics.NumBBInsts[TI.getSuccessor(I)]; - // We don't know which blocks will be eliminated, so use the average size. - Reduction += InlineConstants::InstrCost*Instrs*(NumSucc-1)/NumSucc; - } else if (CallInst *CI = dyn_cast(U)) { - // Turning an indirect call into a direct call is a BIG win - if (CI->getCalledValue() == V) - Reduction += InlineConstants::IndirectCallBonus; - } else if (InvokeInst *II = dyn_cast(U)) { - // Turning an indirect call into a direct call is a BIG win - if (II->getCalledValue() == V) - Reduction += InlineConstants::IndirectCallBonus; - } else { - // Figure out if this instruction will be removed due to simple constant - // propagation. - Instruction &Inst = cast(*U); - - // We can't constant propagate instructions which have effects or - // read memory. - // - // FIXME: It would be nice to capture the fact that a load from a - // pointer-to-constant-global is actually a *really* good thing to zap. - // Unfortunately, we don't know the pointer that may get propagated here, - // so we can't make this decision. - if (Inst.mayReadFromMemory() || Inst.mayHaveSideEffects() || - isa(Inst)) - continue; - - 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 += InlineConstants::InstrCost; - - // And any other instructions that use it which become constants - // themselves. - Reduction += CountCodeReductionForConstant(&Inst); - } - } - } - return Reduction; -} - -// 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. -// -unsigned InlineCostAnalyzer::FunctionInfo:: - CountCodeReductionForAlloca(Value *V) { - if (!V->getType()->isPointerTy()) 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 += InlineConstants::InstrCost; - else if (GetElementPtrInst *GEP = dyn_cast(I)) { - // If the GEP has variable indices, we won't be able to do much with it. - if (GEP->hasAllConstantIndices()) - Reduction += CountCodeReductionForAlloca(GEP); - } else if (BitCastInst *BCI = dyn_cast(I)) { - // Track pointer through bitcasts. - Reduction += CountCodeReductionForAlloca(BCI); - } 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; -} - /// callIsSmall - If a call is likely to lower to a single target instruction, /// or is otherwise deemed small return true. /// TODO: Perhaps calls like memcpy, strcpy, etc? @@ -152,30 +62,40 @@ void CodeMetrics::analyzeBasicBlock(const BasicBlock *BB) { if (isa(II) || isa(II)) { if (isa(II)) continue; // Debug intrinsics don't count as size. - - CallSite CS = CallSite::get(const_cast(&*II)); - + + ImmutableCallSite CS(cast(II)); + // If this function contains a call to setjmp or _setjmp, never inline // it. This is a hack because we depend on the user marking their local // variables as volatile if they are live across a setjmp call, and they // probably won't do this in callers. - if (Function *F = CS.getCalledFunction()) { + if (const Function *F = CS.getCalledFunction()) { + // If a function is both internal and has a single use, then it is + // extremely likely to get inlined in the future (it was probably + // exposed by an interleaved devirtualization pass). + if (F->hasInternalLinkage() && F->hasOneUse()) + ++NumInlineCandidates; + if (F->isDeclaration() && (F->getName() == "setjmp" || F->getName() == "_setjmp")) - NeverInline = true; + callsSetJmp = true; // If this call is to function itself, then the function is recursive. // Inlining it into other functions is a bad idea, because this is // basically just a form of loop peeling, and our metrics aren't useful // for that case. if (F == BB->getParent()) - NeverInline = true; + isRecursive = true; } if (!isa(II) && !callIsSmall(CS.getCalledFunction())) { // Each argument to a call takes on average one instruction to set up. NumInsts += CS.arg_size(); - ++NumCalls; + + // We don't want inline asm to count as a call - that would prevent loop + // unrolling. The argument setup cost is still real, though. + if (!isa(CS.getCalledValue())) + ++NumCalls; } } @@ -216,12 +136,141 @@ void CodeMetrics::analyzeBasicBlock(const BasicBlock *BB) { // jump would jump from the inlined copy of the function into the original // function which is extremely undefined behavior. if (isa(BB->getTerminator())) - NeverInline = true; + containsIndirectBr = true; // Remember NumInsts for this BB. NumBBInsts[BB] = NumInsts - NumInstsBeforeThisBB; } +// CountBonusForConstant - Figure out an approximation for how much per-call +// performance boost we can expect if the specified value is constant. +unsigned CodeMetrics::CountBonusForConstant(Value *V) { + unsigned Bonus = 0; + for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){ + User *U = *UI; + if (CallInst *CI = dyn_cast(U)) { + // Turning an indirect call into a direct call is a BIG win + if (CI->getCalledValue() == V) + Bonus += InlineConstants::IndirectCallBonus; + } + else if (InvokeInst *II = dyn_cast(U)) { + // Turning an indirect call into a direct call is a BIG win + if (II->getCalledValue() == V) + Bonus += InlineConstants::IndirectCallBonus; + } + // FIXME: Eliminating conditional branches and switches should + // also yield a per-call performance boost. + else { + // Figure out the bonuses that wll accrue due to simple constant + // propagation. + Instruction &Inst = cast(*U); + + // We can't constant propagate instructions which have effects or + // read memory. + // + // FIXME: It would be nice to capture the fact that a load from a + // pointer-to-constant-global is actually a *really* good thing to zap. + // Unfortunately, we don't know the pointer that may get propagated here, + // so we can't make this decision. + if (Inst.mayReadFromMemory() || Inst.mayHaveSideEffects() || + isa(Inst)) + continue; + + 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) + Bonus += CountBonusForConstant(&Inst); + } + } + return Bonus; +} + + +// CountCodeReductionForConstant - Figure out an approximation for how many +// instructions will be constant folded if the specified value is constant. +// +unsigned CodeMetrics::CountCodeReductionForConstant(Value *V) { + unsigned Reduction = 0; + for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){ + User *U = *UI; + if (isa(U) || isa(U)) { + // We will be able to eliminate all but one of the successors. + const TerminatorInst &TI = cast(*U); + const unsigned NumSucc = TI.getNumSuccessors(); + unsigned Instrs = 0; + for (unsigned I = 0; I != NumSucc; ++I) + Instrs += NumBBInsts[TI.getSuccessor(I)]; + // We don't know which blocks will be eliminated, so use the average size. + Reduction += InlineConstants::InstrCost*Instrs*(NumSucc-1)/NumSucc; + } else { + // Figure out if this instruction will be removed due to simple constant + // propagation. + Instruction &Inst = cast(*U); + + // We can't constant propagate instructions which have effects or + // read memory. + // + // FIXME: It would be nice to capture the fact that a load from a + // pointer-to-constant-global is actually a *really* good thing to zap. + // Unfortunately, we don't know the pointer that may get propagated here, + // so we can't make this decision. + if (Inst.mayReadFromMemory() || Inst.mayHaveSideEffects() || + isa(Inst)) + continue; + + 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 += InlineConstants::InstrCost; + + // And any other instructions that use it which become constants + // themselves. + Reduction += CountCodeReductionForConstant(&Inst); + } + } + } + return Reduction; +} + +// 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. +// +unsigned CodeMetrics::CountCodeReductionForAlloca(Value *V) { + if (!V->getType()->isPointerTy()) 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 += InlineConstants::InstrCost; + else if (GetElementPtrInst *GEP = dyn_cast(I)) { + // If the GEP has variable indices, we won't be able to do much with it. + if (GEP->hasAllConstantIndices()) + Reduction += CountCodeReductionForAlloca(GEP); + } else if (BitCastInst *BCI = dyn_cast(I)) { + // Track pointer through bitcasts. + Reduction += CountCodeReductionForAlloca(BCI); + } 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 CodeMetrics::analyzeFunction(Function *F) { @@ -243,17 +292,62 @@ void InlineCostAnalyzer::FunctionInfo::analyzeFunction(Function *F) { // Don't bother calculating argument weights if we are never going to inline // the function anyway. - if (Metrics.NeverInline) + if (NeverInline()) return; // 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. ArgumentWeights.reserve(F->arg_size()); for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E; ++I) - ArgumentWeights.push_back(ArgInfo(CountCodeReductionForConstant(I), - CountCodeReductionForAlloca(I))); + ArgumentWeights.push_back(ArgInfo(Metrics.CountCodeReductionForConstant(I), + Metrics.CountCodeReductionForAlloca(I), + Metrics.CountBonusForConstant(I))); } +/// NeverInline - returns true if the function should never be inlined into +/// any caller +bool InlineCostAnalyzer::FunctionInfo::NeverInline() +{ + return (Metrics.callsSetJmp || Metrics.isRecursive || + Metrics.containsIndirectBr); + +} +// getSpecializationBonus - The heuristic used to determine the per-call +// performance boost for using a specialization of Callee with argument +// specializedArgNo replaced by a constant. +int InlineCostAnalyzer::getSpecializationBonus(Function *Callee, + SmallVectorImpl &SpecializedArgNos) +{ + if (Callee->mayBeOverridden()) + return 0; + + int Bonus = 0; + // If this function uses the coldcc calling convention, prefer not to + // specialize it. + if (Callee->getCallingConv() == CallingConv::Cold) + Bonus -= InlineConstants::ColdccPenalty; + + // Get information about the callee. + FunctionInfo *CalleeFI = &CachedFunctionInfo[Callee]; + + // If we haven't calculated this information yet, do so now. + if (CalleeFI->Metrics.NumBlocks == 0) + CalleeFI->analyzeFunction(Callee); + + + for (unsigned i = 0, s = SpecializedArgNos.size(); + i < s; ++i ) + { + Bonus += CalleeFI->ArgumentWeights[SpecializedArgNos[i]].ConstantBonus; + } + // Calls usually take a long time, so they make the specialization gain + // smaller. + Bonus -= CalleeFI->Metrics.NumCalls * InlineConstants::CallPenalty; + + return Bonus; +} + + // getInlineCost - The heuristic used to determine if we should inline the // function call or not. // @@ -311,7 +405,7 @@ InlineCost InlineCostAnalyzer::getInlineCost(CallSite CS, CalleeFI->analyzeFunction(Callee); // If we should never inline this, return a huge cost. - if (CalleeFI->Metrics.NeverInline) + if (CalleeFI->NeverInline()) return InlineCost::getNever(); // FIXME: It would be nice to kill off CalleeFI->NeverInline. Then we @@ -367,7 +461,8 @@ InlineCost InlineCostAnalyzer::getInlineCost(CallSite CS, // away with this information. } else if (isa(I)) { if (ArgNo < CalleeFI->ArgumentWeights.size()) - InlineCost -= CalleeFI->ArgumentWeights[ArgNo].ConstantWeight; + InlineCost -= (CalleeFI->ArgumentWeights[ArgNo].ConstantWeight + + CalleeFI->ArgumentWeights[ArgNo].ConstantBonus); } } @@ -383,6 +478,40 @@ InlineCost InlineCostAnalyzer::getInlineCost(CallSite CS, return llvm::InlineCost::get(InlineCost); } +// getSpecializationCost - The heuristic used to determine the code-size +// impact of creating a specialized version of Callee with argument +// SpecializedArgNo replaced by a constant. +InlineCost InlineCostAnalyzer::getSpecializationCost(Function *Callee, + SmallVectorImpl &SpecializedArgNos) +{ + // Don't specialize functions which can be redefined at link-time to mean + // something else. + if (Callee->mayBeOverridden()) + return llvm::InlineCost::getNever(); + + // Get information about the callee. + FunctionInfo *CalleeFI = &CachedFunctionInfo[Callee]; + + // If we haven't calculated this information yet, do so now. + if (CalleeFI->Metrics.NumBlocks == 0) + CalleeFI->analyzeFunction(Callee); + + int Cost = 0; + + // Look at the orginal size of the callee. Each instruction counts as 5. + Cost += CalleeFI->Metrics.NumInsts * InlineConstants::InstrCost; + + // Offset that with the amount of code that can be constant-folded + // away with the given arguments replaced by constants. + for (SmallVectorImpl::iterator an = SpecializedArgNos.begin(), + ae = SpecializedArgNos.end(); an != ae; ++an) + { + Cost -= CalleeFI->ArgumentWeights[*an].ConstantWeight; + } + + return llvm::InlineCost::get(Cost); +} + // getInlineFudgeFactor - Return a > 1.0 factor if the inliner should use a // higher threshold to determine if the function call should be inlined. float InlineCostAnalyzer::getInlineFudgeFactor(CallSite CS) { @@ -439,10 +568,15 @@ InlineCostAnalyzer::growCachedCostInfo(Function *Caller, Function *Callee) { } // Since CalleeMetrics were already calculated, we know that the CallerMetrics - // reference isn't invalidated: both were in the DenseMap. - CallerMetrics.NeverInline |= CalleeMetrics.NeverInline; + // reference isn't invalidated: both were in the DenseMap. CallerMetrics.usesDynamicAlloca |= CalleeMetrics.usesDynamicAlloca; + // FIXME: If any of these three are true for the callee, the callee was + // not inlined into the caller, so I think they're redundant here. + CallerMetrics.callsSetJmp |= CalleeMetrics.callsSetJmp; + CallerMetrics.isRecursive |= CalleeMetrics.isRecursive; + CallerMetrics.containsIndirectBr |= CalleeMetrics.containsIndirectBr; + CallerMetrics.NumInsts += CalleeMetrics.NumInsts; CallerMetrics.NumBlocks += CalleeMetrics.NumBlocks; CallerMetrics.NumCalls += CalleeMetrics.NumCalls; @@ -455,6 +589,11 @@ InlineCostAnalyzer::growCachedCostInfo(Function *Caller, Function *Callee) { else CallerMetrics.NumInsts = 0; - // We are not updating the argumentweights. We have already determined that + // We are not updating the argument weights. We have already determined that // Caller is a fairly large function, so we accept the loss of precision. } + +/// clear - empty the cache of inline costs +void InlineCostAnalyzer::clear() { + CachedFunctionInfo.clear(); +}