//
// 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 is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// This pass promotes "by reference" arguments to be "by value" arguments. In
// practice, this means looking for internal functions that have pointer
-// arguments. If we can prove, through the use of alias analysis, that an
-// argument is *only* loaded, then we can pass the value into the function
+// arguments. If it can prove, through the use of alias analysis, that an
+// argument is *only* loaded, then it can pass the value into the function
// instead of the address of the value. This can cause recursive simplification
// of code and lead to the elimination of allocas (especially in C++ template
// code like the STL).
//
// This pass also handles aggregate arguments that are passed into a function,
// scalarizing them if the elements of the aggregate are only loaded. Note that
-// we refuse to scalarize aggregates which would require passing in more than
-// three operands to the function, because we don't want to pass thousands of
-// operands for a large array or structure!
+// it refuses to scalarize aggregates which would require passing in more than
+// three operands to the function, because passing thousands of operands for a
+// large array or structure is unprofitable!
//
// Note that this transformation could also be done for arguments that are only
-// stored to (returning the value instead), but we do not currently handle that
-// case. This case would be best handled when and if we start supporting
-// multiple return values from functions.
+// stored to (returning the value instead), but does not currently. This case
+// would be best handled when and if LLVM begins supporting multiple return
+// values from functions.
//
//===----------------------------------------------------------------------===//
#include "llvm/Module.h"
#include "llvm/CallGraphSCCPass.h"
#include "llvm/Instructions.h"
+#include "llvm/ParameterAttributes.h"
#include "llvm/Analysis/AliasAnalysis.h"
#include "llvm/Analysis/CallGraph.h"
#include "llvm/Target/TargetData.h"
#include "llvm/ADT/DepthFirstIterator.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/ADT/StringExtras.h"
-#include <iostream>
+#include "llvm/Support/Compiler.h"
#include <set>
using namespace llvm;
-namespace {
- Statistic<> NumArgumentsPromoted("argpromotion",
- "Number of pointer arguments promoted");
- Statistic<> NumAggregatesPromoted("argpromotion",
- "Number of aggregate arguments promoted");
- Statistic<> NumArgumentsDead("argpromotion",
- "Number of dead pointer args eliminated");
+STATISTIC(NumArgumentsPromoted , "Number of pointer arguments promoted");
+STATISTIC(NumAggregatesPromoted, "Number of aggregate arguments promoted");
+STATISTIC(NumArgumentsDead , "Number of dead pointer args eliminated");
+namespace {
/// ArgPromotion - The 'by reference' to 'by value' argument promotion pass.
///
- struct ArgPromotion : public CallGraphSCCPass {
+ struct VISIBILITY_HIDDEN ArgPromotion : public CallGraphSCCPass {
virtual void getAnalysisUsage(AnalysisUsage &AU) const {
AU.addRequired<AliasAnalysis>();
AU.addRequired<TargetData>();
}
virtual bool runOnSCC(const std::vector<CallGraphNode *> &SCC);
+ static char ID; // Pass identification, replacement for typeid
+ ArgPromotion() : CallGraphSCCPass((intptr_t)&ID) {}
+
private:
bool PromoteArguments(CallGraphNode *CGN);
- bool isSafeToPromoteArgument(Argument *Arg) const;
- Function *DoPromotion(Function *F, std::vector<Argument*> &ArgsToPromote);
+ bool isSafeToPromoteArgument(Argument *Arg, bool isByVal) const;
+ Function *DoPromotion(Function *F,
+ SmallPtrSet<Argument*, 8> &ArgsToPromote);
};
+ char ArgPromotion::ID = 0;
RegisterPass<ArgPromotion> X("argpromotion",
"Promote 'by reference' arguments to scalars");
}
-ModulePass *llvm::createArgumentPromotionPass() {
+Pass *llvm::createArgumentPromotionPass() {
return new ArgPromotion();
}
if (!F || !F->hasInternalLinkage()) return false;
// First check: see if there are any pointer arguments! If not, quick exit.
- std::vector<Argument*> PointerArgs;
- for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E; ++I)
+ SmallVector<std::pair<Argument*, unsigned>, 16> PointerArgs;
+ unsigned ArgNo = 0;
+ for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end();
+ I != E; ++I, ++ArgNo)
if (isa<PointerType>(I->getType()))
- PointerArgs.push_back(I);
+ PointerArgs.push_back(std::pair<Argument*, unsigned>(I, ArgNo));
if (PointerArgs.empty()) return false;
// Second check: make sure that all callers are direct callers. We can't
// Ensure that this call site is CALLING the function, not passing it as
// an argument.
- for (CallSite::arg_iterator AI = CS.arg_begin(), E = CS.arg_end();
- AI != E; ++AI)
- if (*AI == F) return false; // Passing the function address in!
+ if (UI.getOperandNo() != 0)
+ return false;
}
- // Check to see which arguments are promotable. If an argument is not
- // promotable, remove it from the PointerArgs vector.
- for (unsigned i = 0; i != PointerArgs.size(); ++i)
- if (!isSafeToPromoteArgument(PointerArgs[i])) {
- std::swap(PointerArgs[i--], PointerArgs.back());
- PointerArgs.pop_back();
- }
-
+ // Check to see which arguments are promotable. If an argument is promotable,
+ // add it to ArgsToPromote.
+ SmallPtrSet<Argument*, 8> ArgsToPromote;
+ for (unsigned i = 0; i != PointerArgs.size(); ++i) {
+ bool isByVal = F->paramHasAttr(PointerArgs[i].second, ParamAttr::ByVal);
+ if (isSafeToPromoteArgument(PointerArgs[i].first, isByVal))
+ ArgsToPromote.insert(PointerArgs[i].first);
+ }
+
// No promotable pointer arguments.
- if (PointerArgs.empty()) return false;
+ if (ArgsToPromote.empty()) return false;
- // Okay, promote all of the arguments are rewrite the callees!
- Function *NewF = DoPromotion(F, PointerArgs);
+ Function *NewF = DoPromotion(F, ArgsToPromote);
- // Update the call graph to know that the old function is gone.
+ // Update the call graph to know that the function has been transformed.
getAnalysis<CallGraph>().changeFunction(F, NewF);
return true;
}
static bool AllCalleesPassInValidPointerForArgument(Argument *Arg) {
Function *Callee = Arg->getParent();
- unsigned ArgNo = std::distance(Callee->arg_begin(), Function::arg_iterator(Arg));
+ unsigned ArgNo = std::distance(Callee->arg_begin(),
+ Function::arg_iterator(Arg));
// Look at all call sites of the function. At this pointer we know we only
// have direct callees.
return true;
}
-/// AccessOccursOnPath - Returns true if and only if a load or GEP instruction
-/// on Pointer occurs in Path, or in every control-flow path that succeeds it.
-bool AccessOccursOnPath(Value* V, BasicBlock* Start) {
- std::vector<BasicBlock*> Worklist;
- Worklist.push_back(Start);
-
- std::set<BasicBlock*> Visited;
-
- while (!Worklist.empty()) {
- BasicBlock* BB = Worklist.back();
- Worklist.pop_back();
- Visited.insert(BB);
-
- bool ContainsAccess = false;
- for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
- if (isa<LoadInst>(I)) {
- for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end(); OI != OE; ++OI)
- if (*OI == V) {
- ContainsAccess = true;
- break;
- }
- } else if (isa<GetElementPtrInst>(I)) {
- for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end(); OI != OE; ++OI)
- if (*OI == V) {
- ContainsAccess = AccessOccursOnPath(I, I->getParent());
- break;
- }
- }
-
- if (ContainsAccess)
- break;
- }
-
- if (ContainsAccess) continue;
-
- TerminatorInst* TI = BB->getTerminator();
- if (isa<BranchInst>(TI) || isa<SwitchInst>(TI)) {
- for (unsigned i = 0; i < TI->getNumSuccessors(); ++i)
- if (!Visited.count(TI->getSuccessor(i)))
- Worklist.push_back(TI->getSuccessor(i));
- } else {
- return false;
- }
- }
-
- return true;
-}
/// isSafeToPromoteArgument - As you might guess from the name of this method,
/// it checks to see if it is both safe and useful to promote the argument.
/// This method limits promotion of aggregates to only promote up to three
/// elements of the aggregate in order to avoid exploding the number of
/// arguments passed in.
-bool ArgPromotion::isSafeToPromoteArgument(Argument *Arg) const {
+bool ArgPromotion::isSafeToPromoteArgument(Argument *Arg, bool isByVal) const {
// We can only promote this argument if all of the uses are loads, or are GEP
// instructions (with constant indices) that are subsequently loaded.
- bool HasLoadInEntryBlock = false;
+
+ // We can also only promote the load if we can guarantee that it will happen.
+ // Promoting a load causes the load to be unconditionally executed in the
+ // caller, so we can't turn a conditional load into an unconditional load in
+ // general.
+ bool SafeToUnconditionallyLoad = false;
+ if (isByVal) // ByVal arguments are always safe to load from.
+ SafeToUnconditionallyLoad = true;
+
BasicBlock *EntryBlock = Arg->getParent()->begin();
- std::vector<LoadInst*> Loads;
- std::vector<std::vector<ConstantInt*> > GEPIndices;
+ SmallVector<LoadInst*, 16> Loads;
+ std::vector<SmallVector<ConstantInt*, 8> > GEPIndices;
for (Value::use_iterator UI = Arg->use_begin(), E = Arg->use_end();
UI != E; ++UI)
if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
if (LI->isVolatile()) return false; // Don't hack volatile loads
Loads.push_back(LI);
- HasLoadInEntryBlock |= LI->getParent() == EntryBlock;
+
+ // If this load occurs in the entry block, then the pointer is
+ // unconditionally loaded.
+ SafeToUnconditionallyLoad |= LI->getParent() == EntryBlock;
} else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(*UI)) {
if (GEP->use_empty()) {
// Dead GEP's cause trouble later. Just remove them if we run into
// them.
getAnalysis<AliasAnalysis>().deleteValue(GEP);
- GEP->getParent()->getInstList().erase(GEP);
- return isSafeToPromoteArgument(Arg);
+ GEP->eraseFromParent();
+ return isSafeToPromoteArgument(Arg, isByVal);
}
// Ensure that all of the indices are constants.
- std::vector<ConstantInt*> Operands;
+ SmallVector<ConstantInt*, 8> Operands;
for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i)
if (ConstantInt *C = dyn_cast<ConstantInt>(GEP->getOperand(i)))
Operands.push_back(C);
if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
if (LI->isVolatile()) return false; // Don't hack volatile loads
Loads.push_back(LI);
- HasLoadInEntryBlock |= LI->getParent() == EntryBlock;
+
+ // If this load occurs in the entry block, then the pointer is
+ // unconditionally loaded.
+ SafeToUnconditionallyLoad |= LI->getParent() == EntryBlock;
} else {
return false;
}
if (std::find(GEPIndices.begin(), GEPIndices.end(), Operands) ==
GEPIndices.end()) {
if (GEPIndices.size() == 3) {
- DEBUG(std::cerr << "argpromotion disable promoting argument '"
- << Arg->getName() << "' because it would require adding more "
- << "than 3 arguments to the function.\n");
+ DOUT << "argpromotion disable promoting argument '"
+ << Arg->getName() << "' because it would require adding more "
+ << "than 3 arguments to the function.\n";
// We limit aggregate promotion to only promoting up to three elements
// of the aggregate.
return false;
// of the pointer in the entry block of the function) or if we can prove that
// all pointers passed in are always to legal locations (for example, no null
// pointers are passed in, no pointers to free'd memory, etc).
- if (!AccessOccursOnPath(Arg, Arg->getParent()->begin()) &&
+ if (!SafeToUnconditionallyLoad &&
!AllCalleesPassInValidPointerForArgument(Arg))
return false; // Cannot prove that this is safe!!
// it is safe to unconditionally load the pointer. Use alias analysis to
// check to see if the pointer is guaranteed to not be modified from entry of
// the function to each of the load instructions.
- Function &F = *Arg->getParent();
// Because there could be several/many load instructions, remember which
// blocks we know to be transparent to the load.
const PointerType *LoadTy =
cast<PointerType>(Load->getOperand(0)->getType());
- unsigned LoadSize = (unsigned)TD.getTypeSize(LoadTy->getElementType());
+ unsigned LoadSize = (unsigned)TD.getTypeStoreSize(LoadTy->getElementType());
if (AA.canInstructionRangeModify(BB->front(), *Load, Arg, LoadSize))
return false; // Pointer is invalidated!
unsigned idx = 0;
for (; idx < LHS.size() && idx < RHS.size(); ++idx) {
if (LHS[idx] != RHS[idx]) {
- return cast<ConstantInt>(LHS[idx])->getRawValue() <
- cast<ConstantInt>(RHS[idx])->getRawValue();
+ return cast<ConstantInt>(LHS[idx])->getZExtValue() <
+ cast<ConstantInt>(RHS[idx])->getZExtValue();
}
}
/// arguments, and returns the new function. At this point, we know that it's
/// safe to do so.
Function *ArgPromotion::DoPromotion(Function *F,
- std::vector<Argument*> &Args2Prom) {
- std::set<Argument*> ArgsToPromote(Args2Prom.begin(), Args2Prom.end());
+ SmallPtrSet<Argument*, 8> &ArgsToPromote) {
// Start by computing a new prototype for the function, which is the same as
// the old function, but has modified arguments.
// what the new GEP/Load instructions we are inserting look like.
std::map<std::vector<Value*>, LoadInst*> OriginalLoads;
- for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E; ++I)
+ // ParamAttrs - Keep track of the parameter attributes for the arguments
+ // that we are *not* promoting. For the ones that we do promote, the parameter
+ // attributes are lost
+ ParamAttrsVector ParamAttrsVec;
+ const ParamAttrsList *PAL = F->getParamAttrs();
+
+ unsigned index = 1;
+ for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E;
+ ++I, ++index)
if (!ArgsToPromote.count(I)) {
Params.push_back(I->getType());
+ if (PAL) {
+ unsigned attrs = PAL->getParamAttrs(index);
+ if (attrs)
+ ParamAttrsVec.push_back(ParamAttrsWithIndex::get(Params.size(),
+ attrs));
+ }
} else if (I->use_empty()) {
++NumArgumentsDead;
} else {
// Add a parameter to the function for each element passed in.
for (ScalarizeTable::iterator SI = ArgIndices.begin(),
E = ArgIndices.end(); SI != E; ++SI)
- Params.push_back(GetElementPtrInst::getIndexedType(I->getType(), *SI));
+ Params.push_back(GetElementPtrInst::getIndexedType(I->getType(),
+ SI->begin(),
+ SI->end()));
if (ArgIndices.size() == 1 && ArgIndices.begin()->empty())
++NumArgumentsPromoted;
const Type *RetTy = FTy->getReturnType();
+ // Recompute the parameter attributes list based on the new arguments for
+ // the function.
+ if (ParamAttrsVec.empty())
+ PAL = 0;
+ else
+ PAL = ParamAttrsList::get(ParamAttrsVec);
+
// Work around LLVM bug PR56: the CWriter cannot emit varargs functions which
// have zero fixed arguments.
bool ExtraArgHack = false;
if (Params.empty() && FTy->isVarArg()) {
ExtraArgHack = true;
- Params.push_back(Type::IntTy);
+ Params.push_back(Type::Int32Ty);
}
+
+ // Construct the new function type using the new arguments.
FunctionType *NFTy = FunctionType::get(RetTy, Params, FTy->isVarArg());
- // Create the new function body and insert it into the module...
+ // Create the new function body and insert it into the module...
Function *NF = new Function(NFTy, F->getLinkage(), F->getName());
NF->setCallingConv(F->getCallingConv());
+ NF->setParamAttrs(PAL);
+ if (F->hasCollector())
+ NF->setCollector(F->getCollector());
F->getParent()->getFunctionList().insert(F, NF);
// Get the alias analysis information that we need to update to reflect our
Value *V = *AI;
LoadInst *OrigLoad = OriginalLoads[*SI];
if (!SI->empty()) {
- V = new GetElementPtrInst(V, *SI, V->getName()+".idx", Call);
+ V = new GetElementPtrInst(V, SI->begin(), SI->end(),
+ V->getName()+".idx", Call);
AA.copyValue(OrigLoad->getOperand(0), V);
}
Args.push_back(new LoadInst(V, V->getName()+".val", Call));
}
if (ExtraArgHack)
- Args.push_back(Constant::getNullValue(Type::IntTy));
+ Args.push_back(Constant::getNullValue(Type::Int32Ty));
// Push any varargs arguments on the list
for (; AI != CS.arg_end(); ++AI)
Instruction *New;
if (InvokeInst *II = dyn_cast<InvokeInst>(Call)) {
New = new InvokeInst(NF, II->getNormalDest(), II->getUnwindDest(),
- Args, "", Call);
+ Args.begin(), Args.end(), "", Call);
cast<InvokeInst>(New)->setCallingConv(CS.getCallingConv());
+ cast<InvokeInst>(New)->setParamAttrs(PAL);
} else {
- New = new CallInst(NF, Args, "", Call);
+ New = new CallInst(NF, Args.begin(), Args.end(), "", Call);
cast<CallInst>(New)->setCallingConv(CS.getCallingConv());
+ cast<CallInst>(New)->setParamAttrs(PAL);
if (cast<CallInst>(Call)->isTailCall())
cast<CallInst>(New)->setTailCall();
}
if (!Call->use_empty()) {
Call->replaceAllUsesWith(New);
- std::string Name = Call->getName();
- Call->setName("");
- New->setName(Name);
+ New->takeName(Call);
}
// Finally, remove the old call from the program, reducing the use-count of
// F.
- Call->getParent()->getInstList().erase(Call);
+ Call->eraseFromParent();
}
// Since we have now created the new function, splice the body of the old
// Loop over the argument list, transfering uses of the old arguments over to
// the new arguments, also transfering over the names as well.
//
- for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(), I2 = NF->arg_begin();
- I != E; ++I)
+ for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(),
+ I2 = NF->arg_begin(); I != E; ++I)
if (!ArgsToPromote.count(I)) {
// If this is an unmodified argument, move the name and users over to the
// new version.
I->replaceAllUsesWith(I2);
- I2->setName(I->getName());
+ I2->takeName(I);
AA.replaceWithNewValue(I, I2);
++I2;
} else if (I->use_empty()) {
I2->setName(I->getName()+".val");
LI->replaceAllUsesWith(I2);
AA.replaceWithNewValue(LI, I2);
- LI->getParent()->getInstList().erase(LI);
- DEBUG(std::cerr << "*** Promoted load of argument '" << I->getName()
- << "' in function '" << F->getName() << "'\n");
+ LI->eraseFromParent();
+ DOUT << "*** Promoted load of argument '" << I->getName()
+ << "' in function '" << F->getName() << "'\n";
} else {
GetElementPtrInst *GEP = cast<GetElementPtrInst>(I->use_back());
std::vector<Value*> Operands(GEP->op_begin()+1, GEP->op_end());
- unsigned ArgNo = 0;
Function::arg_iterator TheArg = I2;
for (ScalarizeTable::iterator It = ArgIndices.begin();
*It != Operands; ++It, ++TheArg) {
std::string NewName = I->getName();
for (unsigned i = 0, e = Operands.size(); i != e; ++i)
if (ConstantInt *CI = dyn_cast<ConstantInt>(Operands[i]))
- NewName += "."+itostr((int64_t)CI->getRawValue());
+ NewName += "." + CI->getValue().toStringUnsigned(10);
else
NewName += ".x";
TheArg->setName(NewName+".val");
- DEBUG(std::cerr << "*** Promoted agg argument '" << TheArg->getName()
- << "' of function '" << F->getName() << "'\n");
+ DOUT << "*** Promoted agg argument '" << TheArg->getName()
+ << "' of function '" << F->getName() << "'\n";
// All of the uses must be load instructions. Replace them all with
// the argument specified by ArgNo.
LoadInst *L = cast<LoadInst>(GEP->use_back());
L->replaceAllUsesWith(TheArg);
AA.replaceWithNewValue(L, TheArg);
- L->getParent()->getInstList().erase(L);
+ L->eraseFromParent();
}
AA.deleteValue(GEP);
- GEP->getParent()->getInstList().erase(GEP);
+ GEP->eraseFromParent();
}
}
// Notify the alias analysis implementation that we inserted a new argument.
if (ExtraArgHack)
- AA.copyValue(Constant::getNullValue(Type::IntTy), NF->arg_begin());
+ AA.copyValue(Constant::getNullValue(Type::Int32Ty), NF->arg_begin());
// Tell the alias analysis that the old function is about to disappear.
AA.replaceWithNewValue(F, NF);
// Now that the old function is dead, delete it.
- F->getParent()->getFunctionList().erase(F);
+ F->eraseFromParent();
return NF;
}