bool run()
{
ASSERT(m_graph.m_form != SSA);
BlockSet blocksThatNeedInvalidationPoints;
for (BlockIndex blockIndex = m_graph.numBlocks(); blockIndex--;) {
BasicBlock* block = m_graph.block(blockIndex);
if (!block)
continue;
for (unsigned nodeIndex = 0; nodeIndex < block->size(); ++nodeIndex)
handle(nodeIndex, block->at(nodeIndex));
// Note: this assumes that control flow occurs at bytecode instruction boundaries.
if (m_originThatHadFire.isSet()) {
for (unsigned i = block->numSuccessors(); i--;)
blocksThatNeedInvalidationPoints.add(block->successor(i));
}
m_insertionSet.execute(block);
}
for (BlockIndex blockIndex = m_graph.numBlocks(); blockIndex--;) {
BasicBlock* block = m_graph.block(blockIndex);
if (!blocksThatNeedInvalidationPoints.contains(block))
continue;
insertInvalidationCheck(0, block->at(0));
m_insertionSet.execute(block);
}
return true;
}
void ReplayAudioIngest::emit_frame(channel_entry &ch) {
BlockSet bset;
size_t i;
if (ch.fifo->fill_samples( ) < fft_size) {
throw std::runtime_error("cannot emit frame, not enough samples");
}
/* take FFT of ch.fifo->data( ) into ch.current_frame */
fft->compute(ch.current_frame, ch.fifo->data( ));
/* subtract phase of last_frame from phase of current_frame to get output_frame */
for (i = 0; i < fft_size; i++) {
output_frame[i] = std::polar(
std::abs(ch.current_frame[i]),
std::arg(ch.current_frame[i]) - std::arg(ch.last_frame[i])
);
}
/*
* swap last_frame and current_frame pointers
* (this makes current_frame the next last_frame)
*/
std::swap(ch.last_frame, ch.current_frame);
/* write output_frame to buffer */
bset.add_block(REPLAY_PVOC_BLOCK, output_frame, fft_size);
bset.add_block("Debug001", ch.fifo->data( ), fft_size);
ch.buffer->write_blockset(bset);
ch.fifo->pop_samples(fft_hop);
}
timecode_t ReplayBuffer::write_frame(const ReplayFrameData &data) {
BlockSet blkset;
if (data.video_size == 0) {
throw std::runtime_error("Tried to write empty frame");
}
if (data.video_size > 0) {
blkset.add_block(
REPLAY_VIDEO_BLOCK,
data.video_data,
data.video_size
);
}
if (data.thumbnail_size > 0) {
blkset.add_block(
REPLAY_THUMBNAIL_BLOCK,
data.thumbnail_data,
data.thumbnail_size
);
}
if (data.audio != NULL) {
blkset.add_object(REPLAY_AUDIO_BLOCK, *(data.audio));
}
return write_blockset(blkset);
}
static BlockSet getBlocksWithCallsToFuctionsThatObserveSideEffects(
ir::IRKernel& k)
{
BlockSet blocks;
report(" Getting functions that can observe side-effects");
for(auto block = k.cfg()->begin(); block != k.cfg()->end(); ++block)
{
for(auto instruction : block->instructions)
{
auto ptxInstruction = static_cast<ir::PTXInstruction*>(instruction);
// TODO: Check that the target can observe side effects
if(ptxInstruction->isCall())
{
report(" " << ptxInstruction->toString());
blocks.insert(block);
break;
}
}
}
return blocks;
}
static BlockSet getBlocksWithBranchesThatDependOn(
const BlockSet& blocksWithBackwardsBranches,
const InstructionSet& instructionsThatCanObserveSideEffects,
DependenceAnalysis* dependenceAnalysis,
ControlDependenceAnalysis* controlDependenceAnalysis)
{
BlockSet blocksWithDependentBranches;
report(" Getting blocks with branches that can observe side-effects");
for(auto blockWithBranch : blocksWithBackwardsBranches)
{
auto branch = getBranch(blockWithBranch);
if(branch == nullptr) continue;
auto controlDependentInstructions = getControlDependentInstructions(
branch, instructionsThatCanObserveSideEffects,
controlDependenceAnalysis);
for(auto instruction : controlDependentInstructions)
{
if(dependenceAnalysis->dependsOn(instruction, branch))
{
report(" " << blockWithBranch->label());
blocksWithDependentBranches.insert(blockWithBranch);
break;
}
}
}
return blocksWithDependentBranches;
}
Layer Layer::getSubgraphConnectedToTheseOutputs(
const NeuronSet& outputs) const
{
typedef std::set<size_t> BlockSet;
BlockSet blocks;
// TODO: eliminate the reundant inserts
for(auto& output : outputs)
{
size_t block = (output / getOutputBlockingFactor()) % this->blocks();
blocks.insert(block);
}
Layer layer(blocks.size(), getInputBlockingFactor(),
getOutputBlockingFactor(), blockStep());
for(auto& block : blocks)
{
size_t blockIndex = block - *blocks.begin();
layer[blockIndex] = (*this)[block];
layer.at_bias(blockIndex) = at_bias(block);
}
return layer;
}
Udm::Object MatLabUdmChart::distinguish( Udm::Object udmParent ) {
SLSF::State state;
static boost::regex stateflowRegex( "stateflow", boost::regex_constants::perl | boost::regex_constants::icase );
boost::match_results<std::string::const_iterator> results;
SLSF::Subsystem subsystemParent = SLSF::Subsystem::Cast( udmParent );
BlockSet blockSet = subsystemParent.Block_kind_children();
Udm::Object chartParent = udmParent;
for( BlockSet::iterator blsItr = blockSet.begin() ; blsItr != blockSet.end() ; ++blsItr ) {
Block block = *blsItr;
std::string tag( block.Tag() );
if ( regex_search( tag, results, stateflowRegex ) ) {
chartParent = block;
break;
}
}
#if PARADIGM == CyberComposition_PARADIGM
state = SLSF::State::Create( chartParent );
#else
state = SLSF::State::Create( UdmEngine::get_singleton().getTopLevelState() );
SLSF::ConnectorRef connectorRef = SLSF::ConnectorRef::Create( chartParent );
connectorRef.ref() = state;
#endif
return state;
}
BlockSet Drainer::GetBestBlocks() const
{
BlockSet set;
set.reserve(m_RootMacro->m_BestBlocks.size());
for (auto blk : m_RootMacro->m_BestBlocks)
set.push_back(m_Blocks[blk]);
return set;
}
ControlFlowGraph::BlockPointerVector
ControlFlowGraph::reverse_topological_sequence() {
typedef std::set<iterator, BlockSetCompare> BlockSet;
typedef std::queue<iterator> Queue;
report("Creating reverse topological order traversal");
BlockSet visited;
BlockPointerVector sequence;
Queue queue;
queue.push(get_exit_block());
while (sequence.size() != size()) {
if(queue.empty()) {
for (pointer_iterator block = sequence.begin();
block != sequence.end(); ++block) {
for (pointer_iterator pred = (*block)->predecessors.begin();
pred != (*block)->predecessors.end(); ++pred) {
if (visited.count(*pred) == 0) {
queue.push(*pred);
break;
}
}
if(!queue.empty()) {
break;
}
}
if(queue.empty()) break; // The remaining blocks are unreachable
}
iterator current = queue.front();
queue.pop();
if(!visited.insert(current).second) continue;
sequence.push_back(current);
report(" Adding block " << current->label());
for (pointer_iterator block = current->predecessors.begin();
block != current->predecessors.end(); ++block) {
bool noDependencies = true;
for (pointer_iterator successor = (*block)->successors.begin();
successor != (*block)->successors.end(); ++successor) {
if (visited.count(*successor) == 0) {
noDependencies = false;
break;
}
}
if(noDependencies) {
queue.push(*block);
}
}
}
return sequence;
}
void TotalGenerator::Fill(BlockSet &data)
{
std::vector<int_fast8_t> set(data.FullSize(), false);
for (auto i = 0; i < m_Total; i++)
set[i] = true;
std::shuffle(set.begin(), set.end(), m_Random);
for (auto i = 0; i < data.FullSize(); i++)
if (set[i])
data += i;
}
void FindLive(Block *Root) {
BlockList ToInvestigate;
ToInvestigate.push_back(Root);
while (ToInvestigate.size() > 0) {
Block *Curr = ToInvestigate.front();
ToInvestigate.pop_front();
if (Live.find(Curr) != Live.end()) continue;
Live.insert(Curr);
for (BlockBranchMap::iterator iter = Curr->BranchesOut.begin(); iter != Curr->BranchesOut.end(); iter++) {
ToInvestigate.push_back(iter->first);
}
}
}
void TextAutosizer::FingerprintMapper::assertMapsAreConsistent()
{
// For each fingerprint -> block mapping in m_blocksForFingerprint we should have an associated
// map from block -> fingerprint in m_fingerprints.
ReverseFingerprintMap::iterator end = m_blocksForFingerprint.end();
for (ReverseFingerprintMap::iterator fingerprintIt = m_blocksForFingerprint.begin(); fingerprintIt != end; ++fingerprintIt) {
Fingerprint fingerprint = fingerprintIt->key;
BlockSet* blocks = fingerprintIt->value.get();
for (BlockSet::iterator blockIt = blocks->begin(); blockIt != blocks->end(); ++blockIt) {
const LayoutBlock* block = (*blockIt);
ASSERT(m_fingerprints.get(block) == fingerprint);
}
}
}
/**
* Generate side walks
*/
void VBOPmBlocks::generateSideWalk(VBORenderManager* renderManager, BlockSet& blocks) {
for (int i = 0; i < blocks.size(); ++i) {
if (!blocks[i].valid) continue;
BBox3D bbox;
blocks[i].sidewalkContour.getBBox3D(bbox.minPt, bbox.maxPt);
// If the block is too narrow, make it invalid.
if (blocks[i].sidewalkContour.isTooNarrow(8.0f, 18.0f) || blocks[i].sidewalkContour.isTooNarrow(1.0f, 3.0f)) {
blocks[i].valid = false;
continue;
}
// If the block is close or under waterlevel, or on a steep terain, make it invalid.
float min_z = std::numeric_limits<float>::max();
float max_z = 0.0f;
for (int pi = 0; pi < blocks[i].sidewalkContour.contour.size(); ++pi) {
int next_pi = (pi + 1) % blocks[i].sidewalkContour.contour.size();
for (int k = 0; k <= 20; ++k) {
QVector3D pt = blocks[i].sidewalkContour.contour[pi] * (float)(20 - k) * 0.05 + blocks[i].sidewalkContour.contour[next_pi] * (float)k * 0.05;
float z = renderManager->getTerrainHeight(pt.x(), pt.y());
min_z = std::min(min_z, z);
max_z = std::max(max_z, z);
}
//float z = renderManager->getTerrainHeight(blocks[i].sidewalkContour.contour[pi].x(), blocks[i].sidewalkContour.contour[pi].y());
//min_z = std::min(min_z, z);
//max_z = std::max(max_z, z);
}
if (min_z < 40.0f) {
blocks[i].valid = false;
continue;
} else if (max_z - min_z > 20.0f) {
blocks[i].isPark = true;
continue;
}
}
// Compute the block contour (the outer part becomes sidewalks)
for (int i = 0; i < blocks.size(); ++i) {
if (!blocks[i].valid) continue;
//if (blocks[i].isPark) continue;
Loop3D blockContourInset;
float sidewalk_width = G::getFloat("sidewalk_width");
blocks[i].sidewalkContour.computeInset(sidewalk_width, blockContourInset, false);
blocks[i].blockContour.contour = blockContourInset;
//blocks[i].blockContour.getBBox3D(blocks[i].bbox.minPt, blocks[i].bbox.maxPt);
}
}
SILValue
StackAllocationPromoter::getLiveOutValue(BlockSet &PhiBlocks,
SILBasicBlock *StartBB) {
DEBUG(llvm::dbgs() << "*** Searching for a value definition.\n");
// Walk the Dom tree in search of a defining value:
for (DomTreeNode *Node = DT->getNode(StartBB); Node; Node = Node->getIDom()) {
SILBasicBlock *BB = Node->getBlock();
// If there is a store (that must come after the phi), use its value.
BlockToInstMap::iterator it = LastStoreInBlock.find(BB);
if (it != LastStoreInBlock.end())
if (auto *St = dyn_cast_or_null<StoreInst>(it->second)) {
DEBUG(llvm::dbgs() << "*** Found Store def " << *St->getSrc());
return St->getSrc();
}
// If there is a Phi definition in this block:
if (PhiBlocks.count(BB)) {
// Return the dummy instruction that represents the new value that we will
// add to the basic block.
SILValue Phi = BB->getArgument(BB->getNumArguments() - 1);
DEBUG(llvm::dbgs() << "*** Found a dummy Phi def " << *Phi);
return Phi;
}
// Move to the next dominating block.
DEBUG(llvm::dbgs() << "*** Walking up the iDOM.\n");
}
DEBUG(llvm::dbgs() << "*** Could not find a Def. Using Undef.\n");
return SILUndef::get(ASI->getElementType(), ASI->getModule());
}
static bool allPreviousDefinitionsDominateMove(instruction_iterator move,
block_iterator block, BlockSet& visited)
{
if(!visited.insert(block->id()).second) return true;
assert(move->d.size() == 1);
auto destination = *move->d.front().pointer;
// If the value is defined by a phi with multiple sources, then the
// previous definition does not dominate the move
for(auto phi = block->phis().begin(); phi != block->phis().end(); ++phi)
{
if(phi->d == destination)
{
return false;
}
}
// Check all predecessors with the value live out
for(auto predecessor = block->predecessors().begin();
predecessor != block->predecessors().end(); ++predecessor)
{
if((*predecessor)->aliveOut().count(destination) == 0) continue;
if(!allPreviousDefinitionsDominateMove(move, *predecessor, visited))
{
return false;
}
}
return true;
}
static void chaseDownPredecessors(iterator node, Register value,
DataflowGraph* dfg,
dataflow_iterator block, NodeList& nodes,
InstructionToNodeMap& instructionToNodes, BlockSet& visited)
{
if(!visited.insert(block).second) return;
assert(block->aliveIn().count(value) != 0);
for(auto predecessor : block->predecessors())
{
if(predecessor->aliveOut().count(value) == 0) continue;
bool foundAnyDefinitions = false;
// check the body for a definition
for(auto instruction = predecessor->instructions().rbegin();
instruction != predecessor->instructions().rend(); ++instruction)
{
for(auto destination : instruction->d)
{
if(*destination.pointer == value)
{
auto producer = nodes.end();
auto ptx = static_cast<PTXInstruction*>(instruction->i);
auto existingNode = instructionToNodes.find(ptx);
if(existingNode == instructionToNodes.end())
{
producer = nodes.insert(nodes.end(), Node(ptx));
instructionToNodes.insert(
std::make_pair(ptx, producer));
}
else
{
producer = existingNode->second;
}
report(" " << producer->instruction->toString() << " -> "
<< node->instruction->toString());
node->predecessors.push_back(producer);
producer->successors.push_back(node);
foundAnyDefinitions = true;
break;
}
}
}
if(foundAnyDefinitions) continue;
// if no definitions were found, recurse through predecessors
chaseDownPredecessors(node, value, dfg, predecessor, nodes,
instructionToNodes, visited);
}
}
// Create a list of entries from a block. If LimitTo is provided, only results in that set
// will appear
void GetBlocksOut(Block *Source, BlockSet& Entries, BlockSet *LimitTo=NULL) {
for (BlockBranchMap::iterator iter = Source->BranchesOut.begin(); iter != Source->BranchesOut.end(); iter++) {
if (!LimitTo || LimitTo->find(iter->first) != LimitTo->end()) {
Entries.insert(iter->first);
}
}
}
bool isValidFlushLocation(BasicBlock* startingBlock, unsigned index, VirtualRegister operand)
{
// This code is not meant to be fast. We just use it for assertions. If we got liveness wrong,
// this function would return false for a Flush that we insert.
Vector<BasicBlock*, 4> worklist;
BlockSet seen;
auto addPredecessors = [&] (BasicBlock* block) {
for (BasicBlock* predecessor : block->predecessors) {
bool isNewEntry = seen.add(predecessor);
if (isNewEntry)
worklist.append(predecessor);
}
};
auto flushIsDefinitelyInvalid = [&] (BasicBlock* block, unsigned index) {
bool allGood = false;
for (unsigned i = index; i--; ) {
if (block->at(i)->accessesStack(m_graph) && block->at(i)->local() == operand) {
allGood = true;
break;
}
}
if (allGood)
return false;
if (block->predecessors.isEmpty()) {
// This is a root block. We proved we reached here, therefore we can't Flush, as
// it'll make this local live at the start of a root block, which is invalid IR.
return true;
}
addPredecessors(block);
return false;
};
if (flushIsDefinitelyInvalid(startingBlock, index))
return false;
while (!worklist.isEmpty()) {
BasicBlock* block = worklist.takeLast();
if (flushIsDefinitelyInvalid(block, block->size()))
return false;
}
return true;
}
Shape *MakeMultiple(BlockSet &Blocks, BlockSet& Entries, BlockBlockSetMap& IndependentGroups, Shape *Prev, BlockSet &NextEntries) {
PrintDebug("creating multiple block with %d inner groups\n", IndependentGroups.size());
bool Fused = !!(Shape::IsSimple(Prev));
MultipleShape *Multiple = new MultipleShape();
Notice(Multiple);
BlockSet CurrEntries;
for (BlockBlockSetMap::iterator iter = IndependentGroups.begin(); iter != IndependentGroups.end(); iter++) {
Block *CurrEntry = iter->first;
BlockSet &CurrBlocks = iter->second;
PrintDebug(" multiple group with entry %d:\n", CurrEntry->Id);
DebugDump(CurrBlocks, " ");
// Create inner block
CurrEntries.clear();
CurrEntries.insert(CurrEntry);
for (BlockSet::iterator iter = CurrBlocks.begin(); iter != CurrBlocks.end(); iter++) {
Block *CurrInner = *iter;
// Remove the block from the remaining blocks
Blocks.erase(CurrInner);
// Find new next entries and fix branches to them
for (BlockBranchMap::iterator iter = CurrInner->BranchesOut.begin(); iter != CurrInner->BranchesOut.end();) {
Block *CurrTarget = iter->first;
BlockBranchMap::iterator Next = iter;
Next++;
if (!contains(CurrBlocks, CurrTarget)) {
NextEntries.insert(CurrTarget);
Solipsize(CurrTarget, Branch::Break, Multiple, CurrBlocks);
}
iter = Next; // increment carefully because Solipsize can remove us
}
}
Multiple->InnerMap[CurrEntry] = Process(CurrBlocks, CurrEntries, NULL);
// If we are not fused, then our entries will actually be checked
if (!Fused) {
CurrEntry->IsCheckedMultipleEntry = true;
}
}
DebugDump(Blocks, " remaining blocks after multiple:");
// Add entries not handled as next entries, they are deferred
for (BlockSet::iterator iter = Entries.begin(); iter != Entries.end(); iter++) {
Block *Entry = *iter;
if (!contains(IndependentGroups, Entry)) {
NextEntries.insert(Entry);
}
}
return Multiple;
}
Shape *MakeSimple(BlockSet &Blocks, Block *Inner, BlockSet &NextEntries) {
PrintDebug("creating simple block with block #%d\n", Inner->Id);
SimpleShape *Simple = new SimpleShape;
Notice(Simple);
Simple->Inner = Inner;
Inner->Parent = Simple;
if (Blocks.size() > 1) {
Blocks.erase(Inner);
GetBlocksOut(Inner, NextEntries, &Blocks);
BlockSet JustInner;
JustInner.insert(Inner);
for (BlockSet::iterator iter = NextEntries.begin(); iter != NextEntries.end(); iter++) {
Solipsize(*iter, Branch::Direct, Simple, JustInner);
}
}
return Simple;
}
ControlFlowGraph::BlockPointerVector ControlFlowGraph::pre_order_sequence() {
typedef std::unordered_set<iterator> BlockSet;
typedef std::stack<iterator> Stack;
BlockSet visited;
BlockPointerVector sequence;
Stack stack;
if (!empty()) {
stack.push(get_entry_block());
visited.insert(get_entry_block());
}
while (!stack.empty()) {
iterator current = stack.top();
stack.pop();
sequence.push_back(current);
// favor the fallthrough
iterator fallthrough = end();
if (current->has_fallthrough_edge()) {
edge_iterator fallthroughEdge
= sequence.back()->get_fallthrough_edge();
if (visited.insert(fallthroughEdge->tail).second) {
fallthrough = fallthroughEdge->tail;
}
}
for (pointer_iterator block = current->successors.begin();
block != current->successors.end(); ++block) {
if (visited.insert(*block).second) {
stack.push(*block);
}
}
if (fallthrough != end()) {
stack.push(fallthrough);
}
}
return sequence;
}
TextAutosizer::Supercluster* TextAutosizer::getSupercluster(const LayoutBlock* block)
{
Fingerprint fingerprint = m_fingerprintMapper.get(block);
if (!fingerprint)
return nullptr;
BlockSet* roots = m_fingerprintMapper.getTentativeClusterRoots(fingerprint);
if (!roots || roots->size() < 2 || !roots->contains(block))
return nullptr;
SuperclusterMap::AddResult addResult = m_superclusters.add(fingerprint, PassOwnPtr<Supercluster>());
if (!addResult.isNewEntry)
return addResult.storedValue->value.get();
Supercluster* supercluster = new Supercluster(roots);
addResult.storedValue->value = adoptPtr(supercluster);
return supercluster;
}
void StackAllocationPromoter::pruneAllocStackUsage() {
DEBUG(llvm::dbgs() << "*** Pruning : " << *ASI);
BlockSet Blocks;
// Insert all of the blocks that ASI is live in.
for (auto UI = ASI->use_begin(), E = ASI->use_end(); UI != E; ++UI)
Blocks.insert(UI->getUser()->getParent());
// Clear AllocStack state.
LastStoreInBlock.clear();
for (auto Block : Blocks) {
StoreInst *SI = promoteAllocationInBlock(Block);
LastStoreInBlock[Block] = SI;
}
DEBUG(llvm::dbgs() << "*** Finished pruning : " << *ASI);
}
ControlFlowGraph::BlockPointerVector ControlFlowGraph::post_order_sequence() {
typedef std::unordered_set<iterator> BlockSet;
typedef std::stack<iterator> Stack;
report("Creating post order traversal");
BlockSet visited;
BlockPointerVector sequence;
Stack stack;
if (!empty()) {
for (pointer_iterator
block = get_entry_block()->successors.begin();
block != get_entry_block()->successors.end(); ++block) {
if (visited.insert(*block).second) {
stack.push(*block);
}
}
}
while (!stack.empty()) {
iterator current = stack.top();
bool one = false;
for (pointer_iterator block = current->successors.begin();
block != current->successors.end(); ++block) {
if (visited.insert(*block).second) {
stack.push(*block);
one = true;
}
}
if(!one) {
stack.pop();
sequence.push_back(current);
report(" Adding block " << current->label());
}
}
report(" Adding block " << get_entry_block()->label());
sequence.push_back(get_entry_block());
return sequence;
}
void DeadCodeEliminationPass::runOnKernel(ir::IRKernel& k)
{
report("Running dead code elimination on kernel " << k.name);
reportE(REPORT_PTX, k);
Analysis* dfgAnalysis = getAnalysis(Analysis::DataflowGraphAnalysis);
assert(dfgAnalysis != 0);
analysis::DataflowGraph& dfg =
*static_cast<analysis::DataflowGraph*>(dfgAnalysis);
assert(dfg.ssa() != analysis::DataflowGraph::SsaType::None);
BlockSet blocks;
report(" Starting by scanning all basic blocks");
for(iterator block = dfg.begin(); block != dfg.end(); ++block)
{
report(" Queueing up BB_" << block->id());
blocks.insert(block);
}
while(!blocks.empty())
{
iterator block = *blocks.begin();
blocks.erase(blocks.begin());
eliminateDeadInstructions(dfg, blocks, block);
}
report("Finished running dead code elimination on kernel " << k.name);
reportE(REPORT_PTX, k);
}
// If a block has multiple entries but no exits, and it is small enough, it is useful to split it.
// A common example is a C++ function where everything ends up at a final exit block and does some
// RAII cleanup. Without splitting, we will be forced to introduce labelled loops to allow
// reaching the final block
void SplitDeadEnds() {
int TotalCodeSize = 0;
for (BlockSet::iterator iter = Live.begin(); iter != Live.end(); iter++) {
Block *Curr = *iter;
TotalCodeSize += strlen(Curr->Code);
}
for (BlockSet::iterator iter = Live.begin(); iter != Live.end(); iter++) {
Block *Original = *iter;
if (Original->BranchesIn.size() <= 1 || Original->BranchesOut.size() > 0) continue;
if (strlen(Original->Code)*(Original->BranchesIn.size()-1) > TotalCodeSize/5) continue; // if splitting increases raw code size by a significant amount, abort
// Split the node (for simplicity, we replace all the blocks, even though we could have reused the original)
for (BlockBranchMap::iterator iter = Original->BranchesIn.begin(); iter != Original->BranchesIn.end(); iter++) {
Block *Prior = iter->first;
Block *Split = new Block(Original->Code);
Split->BranchesIn[Prior] = new Branch(NULL);
Prior->BranchesOut[Split] = new Branch(Prior->BranchesOut[Original]->Condition, Prior->BranchesOut[Original]->Code);
Prior->BranchesOut.erase(Original);
Parent->AddBlock(Split);
Live.insert(Split);
}
}
}
static BlockSet getBlocksWithBackwardsBranches(CycleAnalysis* cycleAnalysis)
{
auto edges = cycleAnalysis->getAllBackEdges();
report(" Getting blocks with backwards branches");
BlockSet backwardsBranchBlocks;
for(auto& edge : edges)
{
if(edge->type != ir::Edge::Branch) continue;
auto block = edge->head;
if(getBranch(block) == nullptr) continue;
backwardsBranchBlocks.insert(block);
report(" " << block->label());
}
return backwardsBranchBlocks;
}
// Converts/processes all branchings to a specific target
void Solipsize(Block *Target, Branch::FlowType Type, Shape *Ancestor, BlockSet &From) {
PrintDebug("Solipsizing branches into %d\n", Target->Id);
DebugDump(From, " relevant to solipsize: ");
for (BlockSet::iterator iter = Target->BranchesIn.begin(); iter != Target->BranchesIn.end();) {
Block *Prior = *iter;
if (From.find(Prior) == From.end()) {
iter++;
continue;
}
Branch *PriorOut = Prior->BranchesOut[Target];
PriorOut->Ancestor = Ancestor;
PriorOut->Type = Type;
if (MultipleShape *Multiple = Shape::IsMultiple(Ancestor)) {
Multiple->NeedLoop++; // We are breaking out of this Multiple, so need a loop
}
iter++; // carefully increment iter before erasing
Target->BranchesIn.erase(Prior);
Target->ProcessedBranchesIn.insert(Prior);
Prior->BranchesOut.erase(Target);
Prior->ProcessedBranchesOut[Target] = PriorOut;
PrintDebug(" eliminated branch from %d\n", Prior->Id);
}
}
static bool flagHammocksWithSideEffects(SafeRegion& region,
const BlockSet& blocksWithSideEffects)
{
if(region.isLeaf())
{
region.doesNotDependOnSideEffects =
blocksWithSideEffects.count(region.block) == 0;
}
else
{
region.doesNotDependOnSideEffects = true;
for(auto& child : region.children)
{
region.doesNotDependOnSideEffects &=
flagHammocksWithSideEffects(child, blocksWithSideEffects);
}
}
return region.doesNotDependOnSideEffects;
}
static void propagateMoveSourceToUsersInBlock(instruction_iterator move,
instruction_iterator position, block_iterator block, BlockSet& visited)
{
// early exit for visited blocks
if(!visited.insert(block->id()).second) return;
assert(move->d.size() == 1);
assert(move->s.size() == 1);
auto destination = *move->d.front().pointer;
auto moveSource = *move->s.front().pointer;
// We can skip PHIs because the use of a PHI would make the removal illegal
// replace uses in the block
for(; position != block->instructions().end(); ++position)
{
for(auto source = position->s.begin();
source != position->s.end(); ++source)
{
if(*source->pointer == destination)
{
*source->pointer = moveSource;
}
}
}
// replace in successors
for(auto successor = block->successors().begin();
successor != block->successors().end(); ++successor)
{
if((*successor)->aliveIn().count(destination) == 0) continue;
propagateMoveSourceToUsersInBlock(move,
(*successor)->instructions().begin(), *successor, visited);
}
}
