// Calculate the largest possible vregsPassed sets. These are the registers that
// can pass through an MBB live, but may not be live every time. It is assumed
// that all vregsPassed sets are empty before the call.
void MachineVerifier::calcRegsPassed() {
// First push live-out regs to successors' vregsPassed. Remember the MBBs that
// have any vregsPassed.
DenseSet<const MachineBasicBlock*> todo;
for (MachineFunction::const_iterator MFI = MF->begin(), MFE = MF->end();
MFI != MFE; ++MFI) {
const MachineBasicBlock &MBB(*MFI);
BBInfo &MInfo = MBBInfoMap[&MBB];
if (!MInfo.reachable)
continue;
for (MachineBasicBlock::const_succ_iterator SuI = MBB.succ_begin(),
SuE = MBB.succ_end(); SuI != SuE; ++SuI) {
BBInfo &SInfo = MBBInfoMap[*SuI];
if (SInfo.addPassed(MInfo.regsLiveOut))
todo.insert(*SuI);
}
}
// Iteratively push vregsPassed to successors. This will converge to the same
// final state regardless of DenseSet iteration order.
while (!todo.empty()) {
const MachineBasicBlock *MBB = *todo.begin();
todo.erase(MBB);
BBInfo &MInfo = MBBInfoMap[MBB];
for (MachineBasicBlock::const_succ_iterator SuI = MBB->succ_begin(),
SuE = MBB->succ_end(); SuI != SuE; ++SuI) {
if (*SuI == MBB)
continue;
BBInfo &SInfo = MBBInfoMap[*SuI];
if (SInfo.addPassed(MInfo.vregsPassed))
todo.insert(*SuI);
}
}
}
// Calculate the set of virtual registers that must be passed through each basic
// block in order to satisfy the requirements of successor blocks. This is very
// similar to calcRegsPassed, only backwards.
void MachineVerifier::calcRegsRequired() {
// First push live-in regs to predecessors' vregsRequired.
DenseSet<const MachineBasicBlock*> todo;
for (MachineFunction::const_iterator MFI = MF->begin(), MFE = MF->end();
MFI != MFE; ++MFI) {
const MachineBasicBlock &MBB(*MFI);
BBInfo &MInfo = MBBInfoMap[&MBB];
for (MachineBasicBlock::const_pred_iterator PrI = MBB.pred_begin(),
PrE = MBB.pred_end(); PrI != PrE; ++PrI) {
BBInfo &PInfo = MBBInfoMap[*PrI];
if (PInfo.addRequired(MInfo.vregsLiveIn))
todo.insert(*PrI);
}
}
// Iteratively push vregsRequired to predecessors. This will converge to the
// same final state regardless of DenseSet iteration order.
while (!todo.empty()) {
const MachineBasicBlock *MBB = *todo.begin();
todo.erase(MBB);
BBInfo &MInfo = MBBInfoMap[MBB];
for (MachineBasicBlock::const_pred_iterator PrI = MBB->pred_begin(),
PrE = MBB->pred_end(); PrI != PrE; ++PrI) {
if (*PrI == MBB)
continue;
BBInfo &SInfo = MBBInfoMap[*PrI];
if (SInfo.addRequired(MInfo.vregsRequired))
todo.insert(*PrI);
}
}
}
static std::unique_ptr<MemoryBuffer>
ProcessThinLTOModule(Module &TheModule, ModuleSummaryIndex &Index,
StringMap<MemoryBufferRef> &ModuleMap, TargetMachine &TM,
const FunctionImporter::ImportMapTy &ImportList,
const FunctionImporter::ExportSetTy &ExportList,
const DenseSet<GlobalValue::GUID> &GUIDPreservedSymbols,
const GVSummaryMapTy &DefinedGlobals,
const ThinLTOCodeGenerator::CachingOptions &CacheOptions,
bool DisableCodeGen, StringRef SaveTempsDir,
bool Freestanding, unsigned OptLevel, unsigned count) {
// "Benchmark"-like optimization: single-source case
bool SingleModule = (ModuleMap.size() == 1);
if (!SingleModule) {
promoteModule(TheModule, Index);
// Apply summary-based LinkOnce/Weak resolution decisions.
thinLTOResolveWeakForLinkerModule(TheModule, DefinedGlobals);
// Save temps: after promotion.
saveTempBitcode(TheModule, SaveTempsDir, count, ".1.promoted.bc");
}
// Be friendly and don't nuke totally the module when the client didn't
// supply anything to preserve.
if (!ExportList.empty() || !GUIDPreservedSymbols.empty()) {
// Apply summary-based internalization decisions.
thinLTOInternalizeModule(TheModule, DefinedGlobals);
}
// Save internalized bitcode
saveTempBitcode(TheModule, SaveTempsDir, count, ".2.internalized.bc");
if (!SingleModule) {
crossImportIntoModule(TheModule, Index, ModuleMap, ImportList);
// Save temps: after cross-module import.
saveTempBitcode(TheModule, SaveTempsDir, count, ".3.imported.bc");
}
optimizeModule(TheModule, TM, OptLevel, Freestanding);
saveTempBitcode(TheModule, SaveTempsDir, count, ".4.opt.bc");
if (DisableCodeGen) {
// Configured to stop before CodeGen, serialize the bitcode and return.
SmallVector<char, 128> OutputBuffer;
{
raw_svector_ostream OS(OutputBuffer);
ProfileSummaryInfo PSI(TheModule);
auto Index = buildModuleSummaryIndex(TheModule, nullptr, &PSI);
WriteBitcodeToFile(TheModule, OS, true, &Index);
}
return make_unique<SmallVectorMemoryBuffer>(std::move(OutputBuffer));
}
return codegenModule(TheModule, TM);
}
// Find the number of arguments we need to add to the functions.
void CSDataRando::findFunctionArgNodes(const std::vector<const Function *> &Functions) {
std::vector<DSNodeHandle> RootNodes;
for (const Function *F : Functions) {
DSGraph *G = DSA->getDSGraph(*F);
G->getFunctionArgumentsForCall(F, RootNodes);
}
// No additional args to pass.
if (RootNodes.size() == 0) {
return;
}
DenseSet<const DSNode*> MarkedNodes;
for (DSNodeHandle &NH : RootNodes) {
if (DSNode *N = NH.getNode()) {
N->markReachableNodes(MarkedNodes);
}
}
// Remove global nodes from the arg nodes. If we are using the bottom-up
// analysis then if a node is a global node all contexts will use the global map.
for (auto i : GlobalNodes) {
MarkedNodes.erase(i);
}
// Remove any nodes that are marked do not encrypt.
SmallVector<const DSNode*, 8> MarkedNodeWorkList;
for (auto i : MarkedNodes) {
if (i->isDoNotEncryptNode()) {
MarkedNodeWorkList.push_back(i);
}
}
for (auto i : MarkedNodeWorkList) {
MarkedNodes.erase(i);
}
if (MarkedNodes.empty()) {
return;
}
// Create a FuncInfo entry for each of the functions with the arg nodes that
// need to be passed
for (const Function *F : Functions) {
FuncInfo &FI = FunctionInfo[F];
FI.ArgNodes.insert(FI.ArgNodes.end(), MarkedNodes.begin(), MarkedNodes.end());
}
}
static std::unique_ptr<Module>
getModuleForFile(LLVMContext &Context, claimed_file &F, raw_fd_ostream *ApiFile,
StringSet<> &Internalize, StringSet<> &Maybe) {
ld_plugin_input_file File;
if (get_input_file(F.handle, &File) != LDPS_OK)
message(LDPL_FATAL, "Failed to get file information");
if (get_symbols(F.handle, F.syms.size(), &F.syms[0]) != LDPS_OK)
message(LDPL_FATAL, "Failed to get symbol information");
const void *View;
if (get_view(F.handle, &View) != LDPS_OK)
message(LDPL_FATAL, "Failed to get a view of file");
std::unique_ptr<MemoryBuffer> Buffer = MemoryBuffer::getMemBuffer(
StringRef((char *)View, File.filesize), "", false);
if (release_input_file(F.handle) != LDPS_OK)
message(LDPL_FATAL, "Failed to release file information");
ErrorOr<Module *> MOrErr = getLazyBitcodeModule(std::move(Buffer), Context);
if (std::error_code EC = MOrErr.getError())
message(LDPL_FATAL, "Could not read bitcode from file : %s",
EC.message().c_str());
std::unique_ptr<Module> M(MOrErr.get());
SmallPtrSet<GlobalValue *, 8> Used;
collectUsedGlobalVariables(*M, Used, /*CompilerUsed*/ false);
DenseSet<GlobalValue *> Drop;
std::vector<GlobalAlias *> KeptAliases;
for (ld_plugin_symbol &Sym : F.syms) {
ld_plugin_symbol_resolution Resolution =
(ld_plugin_symbol_resolution)Sym.resolution;
if (options::generate_api_file)
*ApiFile << Sym.name << ' ' << getResolutionName(Resolution) << '\n';
GlobalValue *GV = M->getNamedValue(Sym.name);
if (!GV)
continue; // Asm symbol.
if (GV->hasCommonLinkage()) {
// Common linkage is special. There is no single symbol that wins the
// resolution. Instead we have to collect the maximum alignment and size.
// The IR linker does that for us if we just pass it every common GV.
continue;
}
switch (Resolution) {
case LDPR_UNKNOWN:
llvm_unreachable("Unexpected resolution");
case LDPR_RESOLVED_IR:
case LDPR_RESOLVED_EXEC:
case LDPR_RESOLVED_DYN:
case LDPR_UNDEF:
assert(isDeclaration(*GV));
break;
case LDPR_PREVAILING_DEF_IRONLY: {
keepGlobalValue(*GV, KeptAliases);
if (!Used.count(GV)) {
// Since we use the regular lib/Linker, we cannot just internalize GV
// now or it will not be copied to the merged module. Instead we force
// it to be copied and then internalize it.
Internalize.insert(Sym.name);
}
break;
}
case LDPR_PREVAILING_DEF:
keepGlobalValue(*GV, KeptAliases);
break;
case LDPR_PREEMPTED_REG:
case LDPR_PREEMPTED_IR:
Drop.insert(GV);
break;
case LDPR_PREVAILING_DEF_IRONLY_EXP: {
// We can only check for address uses after we merge the modules. The
// reason is that this GV might have a copy in another module
// and in that module the address might be significant, but that
// copy will be LDPR_PREEMPTED_IR.
if (GV->hasLinkOnceODRLinkage())
Maybe.insert(Sym.name);
keepGlobalValue(*GV, KeptAliases);
break;
}
}
free(Sym.name);
free(Sym.comdat_key);
Sym.name = nullptr;
Sym.comdat_key = nullptr;
}
if (!Drop.empty())
// This is horrible. Given how lazy loading is implemented, dropping
// the body while there is a materializer present doesn't work, the
// linker will just read the body back.
M->materializeAllPermanently();
ValueToValueMapTy VM;
LocalValueMaterializer Materializer(Drop);
for (GlobalAlias *GA : KeptAliases) {
// Gold told us to keep GA. It is possible that a GV usied in the aliasee
// expression is being dropped. If that is the case, that GV must be copied.
Constant *Aliasee = GA->getAliasee();
Constant *Replacement = mapConstantToLocalCopy(Aliasee, VM, &Materializer);
if (Aliasee != Replacement)
GA->setAliasee(Replacement);
}
for (auto *GV : Drop)
drop(*GV);
return M;
}
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;
}
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.
// Note: With the migration, we need to recompute this for each 'pass'. Once
// we merge these, we'll do it once before the analysis
DominatorTree DT;
std::vector<CallSite> ParsePointNeeded;
if (EnableBackedgeSafepoints) {
// 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.
FunctionPassManager FPM(F.getParent());
PlaceBackedgeSafepointsImpl *PBS =
new PlaceBackedgeSafepointsImpl(EnableCallSafepoints);
FPM.add(PBS);
// Note: While the analysis pass itself won't modify the IR, LoopSimplify
// (which it depends on) may. i.e. analysis must be recalculated after run
FPM.run(F);
// We preserve dominance information when inserting the poll, otherwise
// we'd have to recalculate this on every insert
DT.recalculate(F);
// Insert a poll at each point the analysis pass identified
for (size_t i = 0; i < PBS->PollLocations.size(); i++) {
// We are inserting a poll, the function is modified
modified = true;
// The poll location must be the terminator of a loop latch block.
TerminatorInst *Term = PBS->PollLocations[i];
std::vector<CallSite> ParsePoints;
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. (Note: This still relies on LoopSimplify.)
DenseSet<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.
DenseSet<BasicBlock *> SplitBackedges;
for (BasicBlock *Header : Headers) {
BasicBlock *NewBB = SplitEdge(Term->getParent(), Header, nullptr);
SplitBackedges.insert(NewBB);
}
DT.recalculate(F);
for (BasicBlock *NewBB : SplitBackedges) {
InsertSafepointPoll(DT, NewBB->getTerminator(), ParsePoints);
NumBackedgeSafepoints++;
}
} else {
// Split the latch block itself, right before the terminator.
InsertSafepointPoll(DT, Term, ParsePoints);
NumBackedgeSafepoints++;
}
// Record the parse points for later use
ParsePointNeeded.insert(ParsePointNeeded.end(), ParsePoints.begin(),
ParsePoints.end());
}
}
if (EnableEntrySafepoints) {
DT.recalculate(F);
Instruction *term = findLocationForEntrySafepoint(F, DT);
if (!term) {
// policy choice not to insert?
} else {
std::vector<CallSite> RuntimeCalls;
InsertSafepointPoll(DT, term, RuntimeCalls);
modified = true;
NumEntrySafepoints++;
ParsePointNeeded.insert(ParsePointNeeded.end(), RuntimeCalls.begin(),
RuntimeCalls.end());
}
}
if (EnableCallSafepoints) {
DT.recalculate(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
DT.recalculate(F); // Needed?
// 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];
Value *GCResult = ReplaceWithStatepoint(CS, nullptr);
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) {
// In case if we inserted result in a different basic block than the
// original safepoint (this can happen for invokes). We need to be sure
// that
// original result value was not used in any of the phi nodes at the
// beginning of basic block with gc result. Because we know that all such
// blocks will have single predecessor we can safely assume that all phi
// nodes have single entry (because of normalizeBBForInvokeSafepoint).
// Just remove them all here.
if (CS.isInvoke()) {
FoldSingleEntryPHINodes(cast<Instruction>(GCResult)->getParent(),
nullptr);
assert(
!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;
}
Result::Sat AttemptSolutionSDP::attempt(const ApproximateSimplex::Solution& sol){
const DenseSet& newBasis = sol.newBasis;
const DenseMap<DeltaRational>& newValues = sol.newValues;
DenseSet needsToBeAdded;
for(DenseSet::const_iterator i = newBasis.begin(), i_end = newBasis.end(); i != i_end; ++i){
ArithVar b = *i;
if(!d_tableau.isBasic(b)){
needsToBeAdded.add(b);
}
}
DenseMap<DeltaRational>::const_iterator nvi = newValues.begin(), nvi_end = newValues.end();
for(; nvi != nvi_end; ++nvi){
ArithVar currentlyNb = *nvi;
if(!d_tableau.isBasic(currentlyNb)){
if(!matchesNewValue(newValues, currentlyNb)){
const DeltaRational& newValue = newValues[currentlyNb];
Trace("arith::updateMany")
<< "updateMany:" << currentlyNb << " "
<< d_variables.getAssignment(currentlyNb) << " to "<< newValue << endl;
d_linEq.update(currentlyNb, newValue);
Assert(d_variables.assignmentIsConsistent(currentlyNb));
}
}
}
d_errorSet.reduceToSignals();
d_errorSet.setSelectionRule(VAR_ORDER);
static int instance = 0;
++instance;
if(processSignals()){
Debug("arith::findModel") << "attemptSolution("<< instance <<") early conflict" << endl;
d_conflictVariables.purge();
return Result::UNSAT;
}else if(d_errorSet.errorEmpty()){
Debug("arith::findModel") << "attemptSolution("<< instance <<") fixed itself" << endl;
return Result::SAT;
}
while(!needsToBeAdded.empty() && !d_errorSet.errorEmpty()){
ArithVar toRemove = ARITHVAR_SENTINEL;
ArithVar toAdd = ARITHVAR_SENTINEL;
DenseSet::const_iterator i = needsToBeAdded.begin(), i_end = needsToBeAdded.end();
for(; toAdd == ARITHVAR_SENTINEL && i != i_end; ++i){
ArithVar v = *i;
Tableau::ColIterator colIter = d_tableau.colIterator(v);
for(; !colIter.atEnd(); ++colIter){
const Tableau::Entry& entry = *colIter;
Assert(entry.getColVar() == v);
ArithVar b = d_tableau.rowIndexToBasic(entry.getRowIndex());
if(!newBasis.isMember(b)){
toAdd = v;
bool favorBOverToRemove =
(toRemove == ARITHVAR_SENTINEL) ||
(matchesNewValue(newValues, toRemove) && !matchesNewValue(newValues, b)) ||
(d_tableau.basicRowLength(toRemove) > d_tableau.basicRowLength(b));
if(favorBOverToRemove){
toRemove = b;
}
}
}
}
Assert(toRemove != ARITHVAR_SENTINEL);
Assert(toAdd != ARITHVAR_SENTINEL);
Trace("arith::forceNewBasis") << toRemove << " " << toAdd << endl;
//Message() << toRemove << " " << toAdd << endl;
d_linEq.pivotAndUpdate(toRemove, toAdd, newValues[toRemove]);
Trace("arith::forceNewBasis") << needsToBeAdded.size() << "to go" << endl;
//Message() << needsToBeAdded.size() << "to go" << endl;
needsToBeAdded.remove(toAdd);
bool conflict = processSignals();
if(conflict){
d_errorSet.reduceToSignals();
d_conflictVariables.purge();
return Result::UNSAT;
}
}
Assert( d_conflictVariables.empty() );
if(d_errorSet.errorEmpty()){
return Result::SAT;
}else{
d_errorSet.reduceToSignals();
return Result::SAT_UNKNOWN;
}
}
bool LoopDependenceAnalysis::isLoopInvariant(const SCEV *S) const {
DenseSet<const Loop*> loops;
getLoops(S, &loops);
return loops.empty();
}
