SetVector<Value*> qdp_jit_vec::get_loads( SetVector<Value*> Vec )
{
SetVector<Value*> loads;
for (Value* V : Vec) {
SetVector<Value*> tmp = get_loads( V );
loads.insert( tmp.begin() , tmp.end() );
}
return loads;
}
bool Linker::linkInMetadata(Module &Src,
DenseMap<unsigned, MDNode *> *ValIDToTempMDMap) {
SetVector<GlobalValue *> ValuesToLink;
if (Mover.move(
Src, ValuesToLink.getArrayRef(),
[this](GlobalValue &GV, IRMover::ValueAdder Add) { assert(false); },
ValIDToTempMDMap, true))
return true;
return false;
}
ScopStmt::ScopStmt(Scop &parent, SmallVectorImpl<unsigned> &Scatter)
: Parent(parent), BB(NULL), IVS(0) {
BaseName = "FinalRead";
// Build iteration domain.
std::string IterationDomainString = "{[i0] : i0 = 0}";
Domain = isl_set_read_from_str(Parent.getCtx(), IterationDomainString.c_str(),
-1);
Domain = isl_set_add_dims(Domain, isl_dim_param, Parent.getNumParams());
Domain = isl_set_set_tuple_name(Domain, getBaseName());
// Build scattering.
unsigned ScatDim = Parent.getMaxLoopDepth() * 2 + 1;
isl_dim *dim = isl_dim_alloc(Parent.getCtx(), Parent.getNumParams(), 1,
ScatDim);
dim = isl_dim_set_tuple_name(dim, isl_dim_out, "scattering");
dim = isl_dim_set_tuple_name(dim, isl_dim_in, getBaseName());
isl_basic_map *bmap = isl_basic_map_universe(isl_dim_copy(dim));
isl_int v;
isl_int_init(v);
isl_constraint *c = isl_equality_alloc(dim);
isl_int_set_si(v, -1);
isl_constraint_set_coefficient(c, isl_dim_out, 0, v);
// TODO: This is incorrect. We should not use a very large number to ensure
// that this statement is executed last.
isl_int_set_si(v, 200000000);
isl_constraint_set_constant(c, v);
bmap = isl_basic_map_add_constraint(bmap, c);
isl_int_clear(v);
Scattering = isl_map_from_basic_map(bmap);
// Build memory accesses, use SetVector to keep the order of memory accesses
// and prevent the same memory access inserted more than once.
SetVector<const Value*> BaseAddressSet;
for (Scop::const_iterator SI = Parent.begin(), SE = Parent.end(); SI != SE;
++SI) {
ScopStmt *Stmt = *SI;
for (MemoryAccessVec::const_iterator I = Stmt->memacc_begin(),
E = Stmt->memacc_end(); I != E; ++I)
BaseAddressSet.insert((*I)->getBaseAddr());
}
for (SetVector<const Value*>::iterator BI = BaseAddressSet.begin(),
BE = BaseAddressSet.end(); BI != BE; ++BI)
MemAccs.push_back(new MemoryAccess(*BI, this));
IsReduction = false;
}
SetVector<Value*> qdp_jit_vec::get_all_linked_stores_from_store( Value* V )
{
bool all=false;
SetVector<Value*> stores;
stores.insert(V);
while(!all) {
SetVector<Value*> new_stores;
new_stores = get_stores( get_loads( stores ) );
all = (new_stores.size() == stores.size());
stores = new_stores;
}
return stores;
}
static void
computeVariableSummary(ModuleSummaryIndex &Index, const GlobalVariable &V,
DenseSet<GlobalValue::GUID> &CantBePromoted) {
SetVector<ValueInfo> RefEdges;
SmallPtrSet<const User *, 8> Visited;
findRefEdges(&V, RefEdges, Visited);
bool NonRenamableLocal = isNonRenamableLocal(V);
GlobalValueSummary::GVFlags Flags(V.getLinkage(), NonRenamableLocal,
/* LiveRoot = */ false);
auto GVarSummary =
llvm::make_unique<GlobalVarSummary>(Flags, RefEdges.takeVector());
if (NonRenamableLocal)
CantBePromoted.insert(V.getGUID());
Index.addGlobalValueSummary(V.getName(), std::move(GVarSummary));
}
/// Determine whether this call has all constant integer arguments (excluding
/// "this") and summarize it to VCalls or ConstVCalls as appropriate.
static void addVCallToSet(DevirtCallSite Call, GlobalValue::GUID Guid,
SetVector<FunctionSummary::VFuncId> &VCalls,
SetVector<FunctionSummary::ConstVCall> &ConstVCalls) {
std::vector<uint64_t> Args;
// Start from the second argument to skip the "this" pointer.
for (auto &Arg : make_range(Call.CS.arg_begin() + 1, Call.CS.arg_end())) {
auto *CI = dyn_cast<ConstantInt>(Arg);
if (!CI || CI->getBitWidth() > 64) {
VCalls.insert({Guid, Call.Offset});
return;
}
Args.push_back(CI->getZExtValue());
}
ConstVCalls.insert({{Guid, Call.Offset}, std::move(Args)});
}
void DeadCodeElimination::processUse(NodeAddr<UseNode*> UA,
SetVector<NodeId> &WorkQ) {
for (NodeAddr<DefNode*> DA : LV.getAllReachingDefs(UA)) {
if (!LiveNodes.count(DA.Id))
WorkQ.insert(DA.Id);
}
}
/// definedInCaller - Return true if the specified value is defined in the
/// function being code extracted, but not in the region being extracted.
/// These values must be passed in as live-ins to the function.
static bool definedInCaller(const SetVector<BasicBlock *> &Blocks, Value *V) {
if (isa<Argument>(V)) return true;
if (Instruction *I = dyn_cast<Instruction>(V))
if (!Blocks.count(I->getParent()))
return true;
return false;
}
// Walk through the operands of a given User via worklist iteration and populate
// the set of GlobalValue references encountered. Invoked either on an
// Instruction or a GlobalVariable (which walks its initializer).
static void findRefEdges(const User *CurUser, SetVector<ValueInfo> &RefEdges,
SmallPtrSet<const User *, 8> &Visited) {
SmallVector<const User *, 32> Worklist;
Worklist.push_back(CurUser);
while (!Worklist.empty()) {
const User *U = Worklist.pop_back_val();
if (!Visited.insert(U).second)
continue;
ImmutableCallSite CS(U);
for (const auto &OI : U->operands()) {
const User *Operand = dyn_cast<User>(OI);
if (!Operand)
continue;
if (isa<BlockAddress>(Operand))
continue;
if (auto *GV = dyn_cast<GlobalValue>(Operand)) {
// We have a reference to a global value. This should be added to
// the reference set unless it is a callee. Callees are handled
// specially by WriteFunction and are added to a separate list.
if (!(CS && CS.isCallee(&OI)))
RefEdges.insert(GV);
continue;
}
Worklist.push_back(Operand);
}
}
}
static SetVector<BasicBlock *> buildExtractionBlockSet(IteratorT BBBegin,
IteratorT BBEnd) {
SetVector<BasicBlock *> Result;
assert(BBBegin != BBEnd);
// Loop over the blocks, adding them to our set-vector, and aborting with an
// empty set if we encounter invalid blocks.
for (IteratorT I = BBBegin, E = BBEnd; I != E; ++I) {
if (!Result.insert(*I))
llvm_unreachable("Repeated basic blocks in extraction input");
if (!isBlockValidForExtraction(**I)) {
Result.clear();
return Result;
}
}
#ifndef NDEBUG
#if LLVM_VERSION_MINOR == 5
for (SetVector<BasicBlock *>::iterator I = std::next(Result.begin()),
#else
for (SetVector<BasicBlock *>::iterator I = llvm::next(Result.begin()),
#endif
E = Result.end();
I != E; ++I)
for (pred_iterator PI = pred_begin(*I), PE = pred_end(*I);
PI != PE; ++PI)
assert(Result.count(*PI) &&
"No blocks in this region may have entries from outside the region"
" except for the first block!");
#endif
return Result;
}
void OMPGenerator::extractValuesFromStruct(SetVector<Value *> OldValues,
Value *Struct,
ValueToValueMapTy &Map) {
for (unsigned i = 0; i < OldValues.size(); i++) {
Value *Address = Builder.CreateStructGEP(Struct, i);
Value *NewValue = Builder.CreateLoad(Address);
Map.insert(std::make_pair(OldValues[i], NewValue));
}
}
void ParallelLoopGenerator::extractValuesFromStruct(
SetVector<Value *> OldValues, Type *Ty, Value *Struct, ValueMapT &Map) {
for (unsigned i = 0; i < OldValues.size(); i++) {
Value *Address = Builder.CreateStructGEP(Ty, Struct, i);
Value *NewValue = Builder.CreateLoad(Address);
NewValue->setName("polly.subfunc.arg." + OldValues[i]->getName());
Map[OldValues[i]] = NewValue;
}
}
void ParallelLoopGenerator::extractValuesFromStruct(
SetVector<Value *> OldValues, Type *Ty, Value *Struct,
ValueToValueMapTy &Map) {
for (unsigned i = 0; i < OldValues.size(); i++) {
Value *Address = Builder.CreateStructGEP(Ty, Struct, i);
Value *NewValue = Builder.CreateLoad(Address);
Map[OldValues[i]] = NewValue;
}
}
TEST(SetVector, EraseTest) {
SetVector<int> S;
S.insert(0);
S.insert(1);
S.insert(2);
auto I = S.erase(std::next(S.begin()));
// Test that the returned iterator is the expected one-after-erase
// and the size/contents is the expected sequence {0, 2}.
EXPECT_EQ(std::next(S.begin()), I);
EXPECT_EQ(2u, S.size());
EXPECT_EQ(0, *S.begin());
EXPECT_EQ(2, *std::next(S.begin()));
}
void DeadCodeElimination::processDef(NodeAddr<DefNode*> DA,
SetVector<NodeId> &WorkQ) {
NodeAddr<InstrNode*> IA = DA.Addr->getOwner(DFG);
for (NodeAddr<UseNode*> UA : IA.Addr->members_if(DFG.IsUse, DFG)) {
if (!LiveNodes.count(UA.Id))
WorkQ.insert(UA.Id);
}
for (NodeAddr<DefNode*> TA : DFG.getRelatedRefs(IA, DA))
LiveNodes.insert(TA.Id);
}
void
CodeGenRegister::addSubRegsPreOrder(SetVector<CodeGenRegister*> &OSet) const {
assert(SubRegsComplete && "Must precompute sub-registers");
std::vector<Record*> Indices = TheDef->getValueAsListOfDefs("SubRegIndices");
for (unsigned i = 0, e = Indices.size(); i != e; ++i) {
CodeGenRegister *SR = SubRegs.find(Indices[i])->second;
if (OSet.insert(SR))
SR->addSubRegsPreOrder(OSet);
}
}
APInt GraphNode::findShortestPath(GraphNode *Dest, SetVector<GraphNode*> Visited) {
if (Dest == this) {
DEBUG(dbgs() << "IneqGraph: Reached: " << *Dest->getValue() << "\n");
return APInt(64, 0, true);
}
DEBUG(dbgs() << "IneqGraph: Node: " << *V << "\n");
if (Visited.count(this)) {
DEBUG(dbgs() << "IneqGraph: Visited\n");
return APInt::getSignedMaxValue(64); //(64, , true);
}
Visited.insert(this);
APInt MayMin = APInt::getSignedMaxValue(64);
for (may_iterator It = may_begin(), E = may_end();
It != E; ++It) {
APInt Total = It->getToEdge()->findShortestPath(Dest, Visited);
if (Total.slt(MayMin))
MayMin = Total + It->getWeight();
}
//DEBUG(dbgs() << "IneqGraph: Node: " << *V << ", MayMin: " << MayMin << "\n");
APInt MustMax = APInt::getSignedMinValue(64);
for (must_iterator It = must_begin(), E = must_end();
It != E; ++It) {
APInt Total = It->getToEdge()->findShortestPath(Dest, Visited);
if (MustMax.slt(Total))
MustMax = Total;
}
// DEBUG(dbgs() << "IneqGraph: Node: " << *V << ", MustMax: " << MustMax << "\n");
if (MustMax.isMinSignedValue()) {
//DEBUG(dbgs() << "IneqGraph: Node: " << *V << ", Distance: " << MayMin << "\n");
DEBUG(dbgs() << "IneqGraph: Ret: " << MayMin << "\n");
return MayMin;
} else {
APInt Min = MustMax.slt(MayMin) ? MustMax : MayMin;
//DEBUG(dbgs() << "IneqGraph: Node: " << *V << ", Distance: " << Min << "\n");
DEBUG(dbgs() << "IneqGraph: Ret: " << Min << "\n");
return Min;
}
}
void DeadCodeElimination::scanInstr(NodeAddr<InstrNode*> IA,
SetVector<NodeId> &WorkQ) {
if (!DFG.IsCode<NodeAttrs::Stmt>(IA))
return;
if (!isLiveInstr(NodeAddr<StmtNode*>(IA).Addr->getCode()))
return;
for (NodeAddr<RefNode*> RA : IA.Addr->members(DFG)) {
if (!LiveNodes.count(RA.Id))
WorkQ.insert(RA.Id);
}
}
bool LiveRangeCalc::isJointlyDominated(const MachineBasicBlock *MBB,
ArrayRef<SlotIndex> Defs,
const SlotIndexes &Indexes) {
const MachineFunction &MF = *MBB->getParent();
BitVector DefBlocks(MF.getNumBlockIDs());
for (SlotIndex I : Defs)
DefBlocks.set(Indexes.getMBBFromIndex(I)->getNumber());
SetVector<unsigned> PredQueue;
PredQueue.insert(MBB->getNumber());
for (unsigned i = 0; i != PredQueue.size(); ++i) {
unsigned BN = PredQueue[i];
if (DefBlocks[BN])
return true;
const MachineBasicBlock *B = MF.getBlockNumbered(BN);
for (const MachineBasicBlock *P : B->predecessors())
PredQueue.insert(P->getNumber());
}
return false;
}
/// \brief Build a set of blocks to extract if the input blocks are viable.
static SetVector<BasicBlock *>
buildExtractionBlockSet(ArrayRef<BasicBlock *> BBs, DominatorTree *DT) {
assert(!BBs.empty() && "The set of blocks to extract must be non-empty");
SetVector<BasicBlock *> Result;
// Loop over the blocks, adding them to our set-vector, and aborting with an
// empty set if we encounter invalid blocks.
for (BasicBlock *BB : BBs) {
// If this block is dead, don't process it.
if (DT && !DT->isReachableFromEntry(BB))
continue;
if (!Result.insert(BB))
llvm_unreachable("Repeated basic blocks in extraction input");
if (!CodeExtractor::isBlockValidForExtraction(*BB)) {
Result.clear();
return Result;
}
}
#ifndef NDEBUG
for (SetVector<BasicBlock *>::iterator I = std::next(Result.begin()),
E = Result.end();
I != E; ++I)
for (pred_iterator PI = pred_begin(*I), PE = pred_end(*I);
PI != PE; ++PI)
assert(Result.count(*PI) &&
"No blocks in this region may have entries from outside the region"
" except for the first block!");
#endif
return Result;
}
bool HexagonDCE::run() {
bool Collected = collect();
if (!Collected)
return false;
const SetVector<NodeId> &DeadNodes = getDeadNodes();
const SetVector<NodeId> &DeadInstrs = getDeadInstrs();
typedef DenseMap<NodeId,NodeId> RefToInstrMap;
RefToInstrMap R2I;
SetVector<NodeId> PartlyDead;
DataFlowGraph &DFG = getDFG();
for (NodeAddr<BlockNode*> BA : DFG.getFunc().Addr->members(DFG)) {
for (auto TA : BA.Addr->members_if(DFG.IsCode<NodeAttrs::Stmt>, DFG)) {
NodeAddr<StmtNode*> SA = TA;
for (NodeAddr<RefNode*> RA : SA.Addr->members(DFG)) {
R2I.insert(std::make_pair(RA.Id, SA.Id));
if (DFG.IsDef(RA) && DeadNodes.count(RA.Id))
if (!DeadInstrs.count(SA.Id))
PartlyDead.insert(SA.Id);
}
}
}
// Nodes to remove.
SetVector<NodeId> Remove = DeadInstrs;
bool Changed = false;
for (NodeId N : PartlyDead) {
auto SA = DFG.addr<StmtNode*>(N);
if (trace())
dbgs() << "Partly dead: " << *SA.Addr->getCode();
Changed |= rewrite(SA, Remove);
}
return erase(Remove) || Changed;
}
SetVector<Value*> qdp_jit_vec::get_stores( Value* V )
{
SetVector<Value*> stores;
SetVector<Value*> to_visit;
to_visit.insert(V);
while (!to_visit.empty()) {
Value* v = to_visit.back();
to_visit.pop_back();
Instruction* vi;
if ((vi = dyn_cast<Instruction>(v))) {
for (Use& U : vi->uses()) {
if (isa<StoreInst>(U.getUser())) {
stores.insert(U.getUser());
} else if (isa<Instruction>(U.getUser())) {
to_visit.insert(U.getUser());
}
}
}
}
return stores;
}
SetVector<Value*> qdp_jit_vec::get_loads( Value* V )
{
SetVector<Value*> loads;
SetVector<Value*> to_visit;
to_visit.insert(V);
while (!to_visit.empty()) {
Value* v = to_visit.back();
to_visit.pop_back();
Instruction* vi;
if ((vi = dyn_cast<Instruction>(v))) {
for (Use& U : vi->operands()) {
if (isa<LoadInst>(U.get())) {
loads.insert(U.get());
} else if (isa<Instruction>(U.get())) {
to_visit.insert(U.get());
}
}
}
}
return loads;
}
SetVector<Value *> ClastStmtCodeGen::getOMPValues(const clast_stmt *Body) {
SetVector<Value *> Values;
// The clast variables
for (CharMapT::iterator I = ClastVars.begin(), E = ClastVars.end(); I != E;
I++)
Values.insert(I->second);
// Find the temporaries that are referenced in the clast statements'
// basic blocks but are not defined by these blocks (e.g., references
// to function arguments or temporaries defined before the start of
// the SCoP).
ParameterVisitor Params;
Params.visit(Body);
for (ParameterVisitor::const_iterator PI = Params.begin(), PE = Params.end();
PI != PE; ++PI) {
Value *V = *PI;
Values.insert(V);
DEBUG(dbgs() << "Adding temporary for OMP copy-in: " << *V << "\n");
}
return Values;
}
// If a linkonce global is present in the MustPreserveSymbols, we need to make
// sure we honor this. To force the compiler to not drop it, we add it to the
// "llvm.compiler.used" global.
void LTOCodeGenerator::preserveDiscardableGVs(
Module &TheModule,
llvm::function_ref<bool(const GlobalValue &)> mustPreserveGV) {
SetVector<Constant *> UsedValuesSet;
if (GlobalVariable *LLVMUsed =
TheModule.getGlobalVariable("llvm.compiler.used")) {
ConstantArray *Inits = cast<ConstantArray>(LLVMUsed->getInitializer());
for (auto &V : Inits->operands())
UsedValuesSet.insert(cast<Constant>(&V));
LLVMUsed->eraseFromParent();
}
llvm::Type *i8PTy = llvm::Type::getInt8PtrTy(TheModule.getContext());
auto mayPreserveGlobal = [&](GlobalValue &GV) {
if (!GV.isDiscardableIfUnused() || GV.isDeclaration())
return;
if (!mustPreserveGV(GV))
return;
if (GV.hasAvailableExternallyLinkage()) {
emitWarning(
(Twine("Linker asked to preserve available_externally global: '") +
GV.getName() + "'").str());
return;
}
if (GV.hasInternalLinkage()) {
emitWarning((Twine("Linker asked to preserve internal global: '") +
GV.getName() + "'").str());
return;
}
UsedValuesSet.insert(ConstantExpr::getBitCast(&GV, i8PTy));
};
for (auto &GV : TheModule)
mayPreserveGlobal(GV);
for (auto &GV : TheModule.globals())
mayPreserveGlobal(GV);
for (auto &GV : TheModule.aliases())
mayPreserveGlobal(GV);
if (UsedValuesSet.empty())
return;
llvm::ArrayType *ATy = llvm::ArrayType::get(i8PTy, UsedValuesSet.size());
auto *LLVMUsed = new llvm::GlobalVariable(
TheModule, ATy, false, llvm::GlobalValue::AppendingLinkage,
llvm::ConstantArray::get(ATy, UsedValuesSet.getArrayRef()),
"llvm.compiler.used");
LLVMUsed->setSection("llvm.metadata");
}
BitVector CodeGenRegBank::computeCoveredRegisters(ArrayRef<Record*> Regs) {
SetVector<const CodeGenRegister*> Set;
// First add Regs with all sub-registers.
for (unsigned i = 0, e = Regs.size(); i != e; ++i) {
CodeGenRegister *Reg = getReg(Regs[i]);
if (Set.insert(Reg))
// Reg is new, add all sub-registers.
// The pre-ordering is not important here.
Reg->addSubRegsPreOrder(Set, *this);
}
// Second, find all super-registers that are completely covered by the set.
for (unsigned i = 0; i != Set.size(); ++i) {
const CodeGenRegister::SuperRegList &SR = Set[i]->getSuperRegs();
for (unsigned j = 0, e = SR.size(); j != e; ++j) {
const CodeGenRegister *Super = SR[j];
if (!Super->CoveredBySubRegs || Set.count(Super))
continue;
// This new super-register is covered by its sub-registers.
bool AllSubsInSet = true;
const CodeGenRegister::SubRegMap &SRM = Super->getSubRegs();
for (CodeGenRegister::SubRegMap::const_iterator I = SRM.begin(),
E = SRM.end(); I != E; ++I)
if (!Set.count(I->second)) {
AllSubsInSet = false;
break;
}
// All sub-registers in Set, add Super as well.
// We will visit Super later to recheck its super-registers.
if (AllSubsInSet)
Set.insert(Super);
}
}
// Convert to BitVector.
BitVector BV(Registers.size() + 1);
for (unsigned i = 0, e = Set.size(); i != e; ++i)
BV.set(Set[i]->EnumValue);
return BV;
}
bool HexagonDCE::rewrite(NodeAddr<InstrNode*> IA, SetVector<NodeId> &Remove) {
if (!getDFG().IsCode<NodeAttrs::Stmt>(IA))
return false;
DataFlowGraph &DFG = getDFG();
MachineInstr *MI = NodeAddr<StmtNode*>(IA).Addr->getCode();
auto &HII = static_cast<const HexagonInstrInfo&>(DFG.getTII());
if (HII.getAddrMode(MI) != HexagonII::PostInc)
return false;
unsigned Opc = MI->getOpcode();
unsigned OpNum, NewOpc;
switch (Opc) {
case Hexagon::L2_loadri_pi:
NewOpc = Hexagon::L2_loadri_io;
OpNum = 1;
break;
case Hexagon::L2_loadrd_pi:
NewOpc = Hexagon::L2_loadrd_io;
OpNum = 1;
break;
case Hexagon::V6_vL32b_pi:
NewOpc = Hexagon::V6_vL32b_ai;
OpNum = 1;
break;
case Hexagon::S2_storeri_pi:
NewOpc = Hexagon::S2_storeri_io;
OpNum = 0;
break;
case Hexagon::S2_storerd_pi:
NewOpc = Hexagon::S2_storerd_io;
OpNum = 0;
break;
case Hexagon::V6_vS32b_pi:
NewOpc = Hexagon::V6_vS32b_ai;
OpNum = 0;
break;
default:
return false;
}
auto IsDead = [this] (NodeAddr<DefNode*> DA) -> bool {
return getDeadNodes().count(DA.Id);
};
NodeList Defs;
MachineOperand &Op = MI->getOperand(OpNum);
for (NodeAddr<DefNode*> DA : IA.Addr->members_if(DFG.IsDef, DFG)) {
if (&DA.Addr->getOp() != &Op)
continue;
Defs = DFG.getRelatedRefs(IA, DA);
if (!std::all_of(Defs.begin(), Defs.end(), IsDead))
return false;
break;
}
// Mark all nodes in Defs for removal.
for (auto D : Defs)
Remove.insert(D.Id);
if (trace())
dbgs() << "Rewriting: " << *MI;
MI->setDesc(HII.get(NewOpc));
MI->getOperand(OpNum+2).setImm(0);
removeOperand(IA, OpNum);
if (trace())
dbgs() << " to: " << *MI;
return true;
}
//
// runMCDesc - Print out MC register descriptions.
//
void
RegisterInfoEmitter::runMCDesc(raw_ostream &OS, CodeGenTarget &Target,
CodeGenRegBank &RegBank) {
EmitSourceFileHeader("MC Register Information", OS);
OS << "\n#ifdef GET_REGINFO_MC_DESC\n";
OS << "#undef GET_REGINFO_MC_DESC\n";
const std::vector<CodeGenRegister*> &Regs = RegBank.getRegisters();
// The lists of sub-registers, super-registers, and overlaps all go in the
// same array. That allows us to share suffixes.
typedef std::vector<const CodeGenRegister*> RegVec;
SmallVector<RegVec, 4> SubRegLists(Regs.size());
SmallVector<RegVec, 4> OverlapLists(Regs.size());
SequenceToOffsetTable<RegVec, CodeGenRegister::Less> RegSeqs;
// Precompute register lists for the SequenceToOffsetTable.
for (unsigned i = 0, e = Regs.size(); i != e; ++i) {
const CodeGenRegister *Reg = Regs[i];
// Compute the ordered sub-register list.
SetVector<const CodeGenRegister*> SR;
Reg->addSubRegsPreOrder(SR, RegBank);
RegVec &SubRegList = SubRegLists[i];
SubRegList.assign(SR.begin(), SR.end());
RegSeqs.add(SubRegList);
// Super-registers are already computed.
const RegVec &SuperRegList = Reg->getSuperRegs();
RegSeqs.add(SuperRegList);
// The list of overlaps doesn't need to have any particular order, except
// Reg itself must be the first element. Pick an ordering that has one of
// the other lists as a suffix.
RegVec &OverlapList = OverlapLists[i];
const RegVec &Suffix = SubRegList.size() > SuperRegList.size() ?
SubRegList : SuperRegList;
CodeGenRegister::Set Omit(Suffix.begin(), Suffix.end());
// First element is Reg itself.
OverlapList.push_back(Reg);
Omit.insert(Reg);
// Any elements not in Suffix.
CodeGenRegister::Set OSet;
Reg->computeOverlaps(OSet, RegBank);
std::set_difference(OSet.begin(), OSet.end(),
Omit.begin(), Omit.end(),
std::back_inserter(OverlapList),
CodeGenRegister::Less());
// Finally, Suffix itself.
OverlapList.insert(OverlapList.end(), Suffix.begin(), Suffix.end());
RegSeqs.add(OverlapList);
}
// Compute the final layout of the sequence table.
RegSeqs.layout();
OS << "namespace llvm {\n\n";
const std::string &TargetName = Target.getName();
// Emit the shared table of register lists.
OS << "extern const uint16_t " << TargetName << "RegLists[] = {\n";
RegSeqs.emit(OS, printRegister);
OS << "};\n\n";
OS << "extern const MCRegisterDesc " << TargetName
<< "RegDesc[] = { // Descriptors\n";
OS << " { \"NOREG\", 0, 0, 0 },\n";
// Emit the register descriptors now.
for (unsigned i = 0, e = Regs.size(); i != e; ++i) {
const CodeGenRegister *Reg = Regs[i];
OS << " { \"" << Reg->getName() << "\", "
<< RegSeqs.get(OverlapLists[i]) << ", "
<< RegSeqs.get(SubRegLists[i]) << ", "
<< RegSeqs.get(Reg->getSuperRegs()) << " },\n";
}
OS << "};\n\n"; // End of register descriptors...
ArrayRef<CodeGenRegisterClass*> RegisterClasses = RegBank.getRegClasses();
// Loop over all of the register classes... emitting each one.
OS << "namespace { // Register classes...\n";
// Emit the register enum value arrays for each RegisterClass
for (unsigned rc = 0, e = RegisterClasses.size(); rc != e; ++rc) {
const CodeGenRegisterClass &RC = *RegisterClasses[rc];
ArrayRef<Record*> Order = RC.getOrder();
// Give the register class a legal C name if it's anonymous.
std::string Name = RC.getName();
// Emit the register list now.
OS << " // " << Name << " Register Class...\n"
<< " const uint16_t " << Name
<< "[] = {\n ";
for (unsigned i = 0, e = Order.size(); i != e; ++i) {
Record *Reg = Order[i];
OS << getQualifiedName(Reg) << ", ";
}
OS << "\n };\n\n";
OS << " // " << Name << " Bit set.\n"
<< " const uint8_t " << Name
<< "Bits[] = {\n ";
BitVectorEmitter BVE;
for (unsigned i = 0, e = Order.size(); i != e; ++i) {
Record *Reg = Order[i];
BVE.add(Target.getRegBank().getReg(Reg)->EnumValue);
}
BVE.print(OS);
OS << "\n };\n\n";
}
OS << "}\n\n";
OS << "extern const MCRegisterClass " << TargetName
<< "MCRegisterClasses[] = {\n";
for (unsigned rc = 0, e = RegisterClasses.size(); rc != e; ++rc) {
const CodeGenRegisterClass &RC = *RegisterClasses[rc];
// Asserts to make sure values will fit in table assuming types from
// MCRegisterInfo.h
assert((RC.SpillSize/8) <= 0xffff && "SpillSize too large.");
assert((RC.SpillAlignment/8) <= 0xffff && "SpillAlignment too large.");
assert(RC.CopyCost >= -128 && RC.CopyCost <= 127 && "Copy cost too large.");
OS << " { " << '\"' << RC.getName() << "\", "
<< RC.getName() << ", " << RC.getName() << "Bits, "
<< RC.getOrder().size() << ", sizeof(" << RC.getName() << "Bits), "
<< RC.getQualifiedName() + "RegClassID" << ", "
<< RC.SpillSize/8 << ", "
<< RC.SpillAlignment/8 << ", "
<< RC.CopyCost << ", "
<< RC.Allocatable << " },\n";
}
OS << "};\n\n";
// Emit the data table for getSubReg().
ArrayRef<CodeGenSubRegIndex*> SubRegIndices = RegBank.getSubRegIndices();
if (SubRegIndices.size()) {
OS << "const uint16_t " << TargetName << "SubRegTable[]["
<< SubRegIndices.size() << "] = {\n";
for (unsigned i = 0, e = Regs.size(); i != e; ++i) {
const CodeGenRegister::SubRegMap &SRM = Regs[i]->getSubRegs();
OS << " /* " << Regs[i]->TheDef->getName() << " */\n";
if (SRM.empty()) {
OS << " {0},\n";
continue;
}
OS << " {";
for (unsigned j = 0, je = SubRegIndices.size(); j != je; ++j) {
// FIXME: We really should keep this to 80 columns...
CodeGenRegister::SubRegMap::const_iterator SubReg =
SRM.find(SubRegIndices[j]);
if (SubReg != SRM.end())
OS << getQualifiedName(SubReg->second->TheDef);
else
OS << "0";
if (j != je - 1)
OS << ", ";
}
OS << "}" << (i != e ? "," : "") << "\n";
}
OS << "};\n\n";
OS << "const uint16_t *get" << TargetName
<< "SubRegTable() {\n return (const uint16_t *)" << TargetName
<< "SubRegTable;\n}\n\n";
}
EmitRegMappingTables(OS, Regs, false);
// Emit Reg encoding table
OS << "extern const uint16_t " << TargetName;
OS << "RegEncodingTable[] = {\n";
// Add entry for NoRegister
OS << " 0,\n";
for (unsigned i = 0, e = Regs.size(); i != e; ++i) {
Record *Reg = Regs[i]->TheDef;
BitsInit *BI = Reg->getValueAsBitsInit("HWEncoding");
uint64_t Value = 0;
for (unsigned b = 0, be = BI->getNumBits(); b != be; ++b) {
if (BitInit *B = dynamic_cast<BitInit*>(BI->getBit(b)))
Value |= (uint64_t)B->getValue() << b;
}
OS << " " << Value << ",\n";
}
OS << "};\n"; // End of HW encoding table
// MCRegisterInfo initialization routine.
OS << "static inline void Init" << TargetName
<< "MCRegisterInfo(MCRegisterInfo *RI, unsigned RA, "
<< "unsigned DwarfFlavour = 0, unsigned EHFlavour = 0) {\n";
OS << " RI->InitMCRegisterInfo(" << TargetName << "RegDesc, "
<< Regs.size()+1 << ", RA, " << TargetName << "MCRegisterClasses, "
<< RegisterClasses.size() << ", " << TargetName << "RegLists, ";
if (SubRegIndices.size() != 0)
OS << "(uint16_t*)" << TargetName << "SubRegTable, "
<< SubRegIndices.size() << ",\n";
else
OS << "NULL, 0,\n";
OS << " " << TargetName << "RegEncodingTable);\n\n";
EmitRegMapping(OS, Regs, false);
OS << "}\n\n";
OS << "} // End llvm namespace \n";
OS << "#endif // GET_REGINFO_MC_DESC\n\n";
}
bool PlaceSafepoints::runOnFunction(Function &F) {
if (F.isDeclaration() || F.empty()) {
// This is a declaration, nothing to do. Must exit early to avoid crash in
// dom tree calculation
return false;
}
if (isGCSafepointPoll(F)) {
// Given we're inlining this inside of safepoint poll insertion, this
// doesn't make any sense. Note that we do make any contained calls
// parseable after we inline a poll.
return false;
}
if (!shouldRewriteFunction(F))
return false;
bool modified = false;
// In various bits below, we rely on the fact that uses are reachable from
// defs. When there are basic blocks unreachable from the entry, dominance
// and reachablity queries return non-sensical results. Thus, we preprocess
// the function to ensure these properties hold.
modified |= removeUnreachableBlocks(F);
// STEP 1 - Insert the safepoint polling locations. We do not need to
// actually insert parse points yet. That will be done for all polls and
// calls in a single pass.
DominatorTree DT;
DT.recalculate(F);
SmallVector<Instruction *, 16> PollsNeeded;
std::vector<CallSite> ParsePointNeeded;
if (enableBackedgeSafepoints(F)) {
// Construct a pass manager to run the LoopPass backedge logic. We
// need the pass manager to handle scheduling all the loop passes
// appropriately. Doing this by hand is painful and just not worth messing
// with for the moment.
legacy::FunctionPassManager FPM(F.getParent());
bool CanAssumeCallSafepoints = enableCallSafepoints(F);
PlaceBackedgeSafepointsImpl *PBS =
new PlaceBackedgeSafepointsImpl(CanAssumeCallSafepoints);
FPM.add(PBS);
FPM.run(F);
// We preserve dominance information when inserting the poll, otherwise
// we'd have to recalculate this on every insert
DT.recalculate(F);
auto &PollLocations = PBS->PollLocations;
auto OrderByBBName = [](Instruction *a, Instruction *b) {
return a->getParent()->getName() < b->getParent()->getName();
};
// We need the order of list to be stable so that naming ends up stable
// when we split edges. This makes test cases much easier to write.
std::sort(PollLocations.begin(), PollLocations.end(), OrderByBBName);
// We can sometimes end up with duplicate poll locations. This happens if
// a single loop is visited more than once. The fact this happens seems
// wrong, but it does happen for the split-backedge.ll test case.
PollLocations.erase(std::unique(PollLocations.begin(),
PollLocations.end()),
PollLocations.end());
// Insert a poll at each point the analysis pass identified
// The poll location must be the terminator of a loop latch block.
for (TerminatorInst *Term : PollLocations) {
// We are inserting a poll, the function is modified
modified = true;
if (SplitBackedge) {
// Split the backedge of the loop and insert the poll within that new
// basic block. This creates a loop with two latches per original
// latch (which is non-ideal), but this appears to be easier to
// optimize in practice than inserting the poll immediately before the
// latch test.
// Since this is a latch, at least one of the successors must dominate
// it. Its possible that we have a) duplicate edges to the same header
// and b) edges to distinct loop headers. We need to insert pools on
// each.
SetVector<BasicBlock *> Headers;
for (unsigned i = 0; i < Term->getNumSuccessors(); i++) {
BasicBlock *Succ = Term->getSuccessor(i);
if (DT.dominates(Succ, Term->getParent())) {
Headers.insert(Succ);
}
}
assert(!Headers.empty() && "poll location is not a loop latch?");
// The split loop structure here is so that we only need to recalculate
// the dominator tree once. Alternatively, we could just keep it up to
// date and use a more natural merged loop.
SetVector<BasicBlock *> SplitBackedges;
for (BasicBlock *Header : Headers) {
BasicBlock *NewBB = SplitEdge(Term->getParent(), Header, &DT);
PollsNeeded.push_back(NewBB->getTerminator());
NumBackedgeSafepoints++;
}
} else {
// Split the latch block itself, right before the terminator.
PollsNeeded.push_back(Term);
NumBackedgeSafepoints++;
}
}
}
if (enableEntrySafepoints(F)) {
Instruction *Location = findLocationForEntrySafepoint(F, DT);
if (!Location) {
// policy choice not to insert?
} else {
PollsNeeded.push_back(Location);
modified = true;
NumEntrySafepoints++;
}
}
// Now that we've identified all the needed safepoint poll locations, insert
// safepoint polls themselves.
for (Instruction *PollLocation : PollsNeeded) {
std::vector<CallSite> RuntimeCalls;
InsertSafepointPoll(PollLocation, RuntimeCalls);
ParsePointNeeded.insert(ParsePointNeeded.end(), RuntimeCalls.begin(),
RuntimeCalls.end());
}
// If we've been asked to not wrap the calls with gc.statepoint, then we're
// done. In the near future, this option will be "constant folded" to true,
// and the code below that deals with insert gc.statepoint calls will be
// removed. Wrapping potentially safepointing calls in gc.statepoint will
// then become the responsibility of the RewriteStatepointsForGC pass.
if (NoStatepoints)
return modified;
PollsNeeded.clear(); // make sure we don't accidentally use
// The dominator tree has been invalidated by the inlining performed in the
// above loop. TODO: Teach the inliner how to update the dom tree?
DT.recalculate(F);
if (enableCallSafepoints(F)) {
std::vector<CallSite> Calls;
findCallSafepoints(F, Calls);
NumCallSafepoints += Calls.size();
ParsePointNeeded.insert(ParsePointNeeded.end(), Calls.begin(), Calls.end());
}
// Unique the vectors since we can end up with duplicates if we scan the call
// site for call safepoints after we add it for entry or backedge. The
// only reason we need tracking at all is that some functions might have
// polls but not call safepoints and thus we might miss marking the runtime
// calls for the polls. (This is useful in test cases!)
unique_unsorted(ParsePointNeeded);
// Any parse point (no matter what source) will be handled here
// We're about to start modifying the function
if (!ParsePointNeeded.empty())
modified = true;
// Now run through and insert the safepoints, but do _NOT_ update or remove
// any existing uses. We have references to live variables that need to
// survive to the last iteration of this loop.
std::vector<Value *> Results;
Results.reserve(ParsePointNeeded.size());
for (size_t i = 0; i < ParsePointNeeded.size(); i++) {
CallSite &CS = ParsePointNeeded[i];
// For invoke statepoints we need to remove all phi nodes at the normal
// destination block.
// Reason for this is that we can place gc_result only after last phi node
// in basic block. We will get malformed code after RAUW for the
// gc_result if one of this phi nodes uses result from the invoke.
if (InvokeInst *Invoke = dyn_cast<InvokeInst>(CS.getInstruction())) {
normalizeForInvokeSafepoint(Invoke->getNormalDest(),
Invoke->getParent());
}
Value *GCResult = ReplaceWithStatepoint(CS);
Results.push_back(GCResult);
}
assert(Results.size() == ParsePointNeeded.size());
// Adjust all users of the old call sites to use the new ones instead
for (size_t i = 0; i < ParsePointNeeded.size(); i++) {
CallSite &CS = ParsePointNeeded[i];
Value *GCResult = Results[i];
if (GCResult) {
// Can not RAUW for the invoke gc result in case of phi nodes preset.
assert(CS.isCall() || !isa<PHINode>(cast<Instruction>(GCResult)->getParent()->begin()));
// Replace all uses with the new call
CS.getInstruction()->replaceAllUsesWith(GCResult);
}
// Now that we've handled all uses, remove the original call itself
// Note: The insert point can't be the deleted instruction!
CS.getInstruction()->eraseFromParent();
}
return modified;
}
bool AAEval::runOnFunction(Function &F) {
AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
SetVector<Value *> Pointers;
SetVector<CallSite> CallSites;
SetVector<Value *> Loads;
SetVector<Value *> Stores;
for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I)
if (I->getType()->isPointerTy()) // Add all pointer arguments.
Pointers.insert(I);
for (inst_iterator I = inst_begin(F), E = inst_end(F); I != E; ++I) {
if (I->getType()->isPointerTy()) // Add all pointer instructions.
Pointers.insert(&*I);
if (EvalTBAA && isa<LoadInst>(&*I))
Loads.insert(&*I);
if (EvalTBAA && isa<StoreInst>(&*I))
Stores.insert(&*I);
Instruction &Inst = *I;
if (CallSite CS = cast<Value>(&Inst)) {
Value *Callee = CS.getCalledValue();
// Skip actual functions for direct function calls.
if (!isa<Function>(Callee) && isInterestingPointer(Callee))
Pointers.insert(Callee);
// Consider formals.
for (CallSite::arg_iterator AI = CS.arg_begin(), AE = CS.arg_end();
AI != AE; ++AI)
if (isInterestingPointer(*AI))
Pointers.insert(*AI);
CallSites.insert(CS);
} else {
// Consider all operands.
for (Instruction::op_iterator OI = Inst.op_begin(), OE = Inst.op_end();
OI != OE; ++OI)
if (isInterestingPointer(*OI))
Pointers.insert(*OI);
}
}
if (PrintNoAlias || PrintMayAlias || PrintPartialAlias || PrintMustAlias ||
PrintNoModRef || PrintMod || PrintRef || PrintModRef)
errs() << "Function: " << F.getName() << ": " << Pointers.size()
<< " pointers, " << CallSites.size() << " call sites\n";
// iterate over the worklist, and run the full (n^2)/2 disambiguations
for (SetVector<Value *>::iterator I1 = Pointers.begin(), E = Pointers.end();
I1 != E; ++I1) {
uint64_t I1Size = AliasAnalysis::UnknownSize;
Type *I1ElTy = cast<PointerType>((*I1)->getType())->getElementType();
if (I1ElTy->isSized()) I1Size = AA.getTypeStoreSize(I1ElTy);
for (SetVector<Value *>::iterator I2 = Pointers.begin(); I2 != I1; ++I2) {
uint64_t I2Size = AliasAnalysis::UnknownSize;
Type *I2ElTy =cast<PointerType>((*I2)->getType())->getElementType();
if (I2ElTy->isSized()) I2Size = AA.getTypeStoreSize(I2ElTy);
switch (AA.alias(*I1, I1Size, *I2, I2Size)) {
case AliasAnalysis::NoAlias:
PrintResults("NoAlias", PrintNoAlias, *I1, *I2, F.getParent());
++NoAlias; break;
case AliasAnalysis::MayAlias:
PrintResults("MayAlias", PrintMayAlias, *I1, *I2, F.getParent());
++MayAlias; break;
case AliasAnalysis::PartialAlias:
PrintResults("PartialAlias", PrintPartialAlias, *I1, *I2,
F.getParent());
++PartialAlias; break;
case AliasAnalysis::MustAlias:
PrintResults("MustAlias", PrintMustAlias, *I1, *I2, F.getParent());
++MustAlias; break;
}
}
}
if (EvalTBAA) {
// iterate over all pairs of load, store
for (SetVector<Value *>::iterator I1 = Loads.begin(), E = Loads.end();
I1 != E; ++I1) {
for (SetVector<Value *>::iterator I2 = Stores.begin(), E2 = Stores.end();
I2 != E2; ++I2) {
switch (AA.alias(AA.getLocation(cast<LoadInst>(*I1)),
AA.getLocation(cast<StoreInst>(*I2)))) {
case AliasAnalysis::NoAlias:
PrintLoadStoreResults("NoAlias", PrintNoAlias, *I1, *I2,
F.getParent());
++NoAlias; break;
case AliasAnalysis::MayAlias:
PrintLoadStoreResults("MayAlias", PrintMayAlias, *I1, *I2,
F.getParent());
++MayAlias; break;
case AliasAnalysis::PartialAlias:
PrintLoadStoreResults("PartialAlias", PrintPartialAlias, *I1, *I2,
F.getParent());
++PartialAlias; break;
case AliasAnalysis::MustAlias:
PrintLoadStoreResults("MustAlias", PrintMustAlias, *I1, *I2,
F.getParent());
++MustAlias; break;
}
}
}
// iterate over all pairs of store, store
for (SetVector<Value *>::iterator I1 = Stores.begin(), E = Stores.end();
I1 != E; ++I1) {
for (SetVector<Value *>::iterator I2 = Stores.begin(); I2 != I1; ++I2) {
switch (AA.alias(AA.getLocation(cast<StoreInst>(*I1)),
AA.getLocation(cast<StoreInst>(*I2)))) {
case AliasAnalysis::NoAlias:
PrintLoadStoreResults("NoAlias", PrintNoAlias, *I1, *I2,
F.getParent());
++NoAlias; break;
case AliasAnalysis::MayAlias:
PrintLoadStoreResults("MayAlias", PrintMayAlias, *I1, *I2,
F.getParent());
++MayAlias; break;
case AliasAnalysis::PartialAlias:
PrintLoadStoreResults("PartialAlias", PrintPartialAlias, *I1, *I2,
F.getParent());
++PartialAlias; break;
case AliasAnalysis::MustAlias:
PrintLoadStoreResults("MustAlias", PrintMustAlias, *I1, *I2,
F.getParent());
++MustAlias; break;
}
}
}
}
// Mod/ref alias analysis: compare all pairs of calls and values
for (SetVector<CallSite>::iterator C = CallSites.begin(),
Ce = CallSites.end(); C != Ce; ++C) {
Instruction *I = C->getInstruction();
for (SetVector<Value *>::iterator V = Pointers.begin(), Ve = Pointers.end();
V != Ve; ++V) {
uint64_t Size = AliasAnalysis::UnknownSize;
Type *ElTy = cast<PointerType>((*V)->getType())->getElementType();
if (ElTy->isSized()) Size = AA.getTypeStoreSize(ElTy);
switch (AA.getModRefInfo(*C, *V, Size)) {
case AliasAnalysis::NoModRef:
PrintModRefResults("NoModRef", PrintNoModRef, I, *V, F.getParent());
++NoModRef; break;
case AliasAnalysis::Mod:
PrintModRefResults("Just Mod", PrintMod, I, *V, F.getParent());
++Mod; break;
case AliasAnalysis::Ref:
PrintModRefResults("Just Ref", PrintRef, I, *V, F.getParent());
++Ref; break;
case AliasAnalysis::ModRef:
PrintModRefResults("Both ModRef", PrintModRef, I, *V, F.getParent());
++ModRef; break;
}
}
}
// Mod/ref alias analysis: compare all pairs of calls
for (SetVector<CallSite>::iterator C = CallSites.begin(),
Ce = CallSites.end(); C != Ce; ++C) {
for (SetVector<CallSite>::iterator D = CallSites.begin(); D != Ce; ++D) {
if (D == C)
continue;
switch (AA.getModRefInfo(*C, *D)) {
case AliasAnalysis::NoModRef:
PrintModRefResults("NoModRef", PrintNoModRef, *C, *D, F.getParent());
++NoModRef; break;
case AliasAnalysis::Mod:
PrintModRefResults("Just Mod", PrintMod, *C, *D, F.getParent());
++Mod; break;
case AliasAnalysis::Ref:
PrintModRefResults("Just Ref", PrintRef, *C, *D, F.getParent());
++Ref; break;
case AliasAnalysis::ModRef:
PrintModRefResults("Both ModRef", PrintModRef, *C, *D, F.getParent());
++ModRef; break;
}
}
}
return false;
}
