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);
}
Shape *MakeLoop(BlockSet &Blocks, BlockSet& Entries, BlockSet &NextEntries) {
// Find the inner blocks in this loop. Proceed backwards from the entries until
// you reach a seen block, collecting as you go.
BlockSet InnerBlocks;
BlockSet Queue = Entries;
while (Queue.size() > 0) {
Block *Curr = *(Queue.begin());
Queue.erase(Queue.begin());
if (InnerBlocks.find(Curr) == InnerBlocks.end()) {
// This element is new, mark it as inner and remove from outer
InnerBlocks.insert(Curr);
Blocks.erase(Curr);
// Add the elements prior to it
for (BlockBranchMap::iterator iter = Curr->BranchesIn.begin(); iter != Curr->BranchesIn.end(); iter++) {
Queue.insert(iter->first);
}
}
}
assert(InnerBlocks.size() > 0);
for (BlockSet::iterator iter = InnerBlocks.begin(); iter != InnerBlocks.end(); iter++) {
Block *Curr = *iter;
for (BlockBranchMap::iterator iter = Curr->BranchesOut.begin(); iter != Curr->BranchesOut.end(); iter++) {
Block *Possible = iter->first;
if (InnerBlocks.find(Possible) == InnerBlocks.end() &&
NextEntries.find(Possible) == NextEntries.find(Possible)) {
NextEntries.insert(Possible);
}
}
}
PrintDebug("creating loop block:\n");
DebugDump(InnerBlocks, " inner blocks:");
DebugDump(Entries, " inner entries:");
DebugDump(Blocks, " outer blocks:");
DebugDump(NextEntries, " outer entries:");
// TODO: Optionally hoist additional blocks into the loop
LoopShape *Loop = new LoopShape();
Notice(Loop);
// Solipsize the loop, replacing with break/continue and marking branches as Processed (will not affect later calculations)
// A. Branches to the loop entries become a continue to this shape
for (BlockSet::iterator iter = Entries.begin(); iter != Entries.end(); iter++) {
Solipsize(*iter, Branch::Continue, Loop, InnerBlocks);
}
// B. Branches to outside the loop (a next entry) become breaks on this shape
for (BlockSet::iterator iter = NextEntries.begin(); iter != NextEntries.end(); iter++) {
Solipsize(*iter, Branch::Break, Loop, InnerBlocks);
}
// Finish up
Shape *Inner = Process(InnerBlocks, Entries, NULL);
Loop->Inner = Inner;
return Loop;
}
// 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() {
unsigned TotalCodeSize = 0;
for (BlockSet::iterator iter = Live.begin(); iter != Live.end(); iter++) {
Block *Curr = *iter;
TotalCodeSize += strlen(Curr->Code);
}
BlockSet Splits;
BlockSet Removed;
//DebugDump(Live, "before");
for (BlockSet::iterator iter = Live.begin(); iter != Live.end(); iter++) {
Block *Original = *iter;
if (Original->BranchesIn.size() <= 1 || Original->BranchesOut.size() > 0) continue; // only dead ends, for now
if (contains(Original->BranchesOut, Original)) continue; // cannot split a looping node
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)
PrintDebug("Splitting block %d\n", Original->Id);
for (BlockSet::iterator iter = Original->BranchesIn.begin(); iter != Original->BranchesIn.end(); iter++) {
Block *Prior = *iter;
Block *Split = new Block(Original->Code, Original->BranchVar);
Parent->AddBlock(Split);
PrintDebug(" to %d\n", Split->Id);
Split->BranchesIn.insert(Prior);
Branch *Details = Prior->BranchesOut[Original];
Prior->BranchesOut[Split] = new Branch(Details->Condition, Details->Code);
Prior->BranchesOut.erase(Original);
for (BlockBranchMap::iterator iter = Original->BranchesOut.begin(); iter != Original->BranchesOut.end(); iter++) {
Block *Post = iter->first;
Branch *Details = iter->second;
Split->BranchesOut[Post] = new Branch(Details->Condition, Details->Code);
Post->BranchesIn.insert(Split);
}
Splits.insert(Split);
Removed.insert(Original);
}
for (BlockBranchMap::iterator iter = Original->BranchesOut.begin(); iter != Original->BranchesOut.end(); iter++) {
Block *Post = iter->first;
Post->BranchesIn.erase(Original);
}
//DebugDump(Live, "mid");
}
for (BlockSet::iterator iter = Splits.begin(); iter != Splits.end(); iter++) {
Live.insert(*iter);
}
for (BlockSet::iterator iter = Removed.begin(); iter != Removed.end(); iter++) {
Live.erase(*iter);
}
//DebugDump(Live, "after");
}
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 (CurrBlocks.find(CurrTarget) == CurrBlocks.end()) {
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 (IndependentGroups.find(Entry) == IndependentGroups.end()) {
NextEntries.insert(Entry);
}
}
return Multiple;
}
void ConstantPropagationPass::runOnKernel(ir::IRKernel& k)
{
report("Running constant propagation on kernel " << k.name);
Analysis* dfgAnalysis = getAnalysis("DataflowGraphAnalysis");
assert(dfgAnalysis != 0);
analysis::DataflowGraph& dfg =
*static_cast<analysis::DataflowGraph*>(dfgAnalysis);
dfg.convertToSSAType(analysis::DataflowGraph::Minimal);
assert(dfg.ssa() == analysis::DataflowGraph::Minimal);
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());
eliminateRedundantInstructions(dfg, blocks, block);
}
report("Finished running constant propagation on kernel " << k.name);
reportE(REPORT_PTX, k);
}
ControlFlowGraph::ConstBlockPointerVector
ControlFlowGraph::executable_sequence() const {
typedef std::unordered_set<const_iterator> BlockSet;
ConstBlockPointerVector sequence;
BlockSet unscheduled;
for(const_iterator i = begin(); i != end(); ++i) {
unscheduled.insert(i);
}
report("Getting executable sequence.");
sequence.push_back(get_entry_block());
unscheduled.erase(get_entry_block());
report(" added " << get_entry_block()->label());
while (!unscheduled.empty()) {
if (sequence.back()->has_fallthrough_edge()) {
const_edge_iterator fallthroughEdge
= sequence.back()->get_fallthrough_edge();
sequence.push_back(fallthroughEdge->tail);
unscheduled.erase(fallthroughEdge->tail);
}
else {
// find a new block, favor branch targets over random blocks
const_iterator next = *unscheduled.begin();
for(const_edge_pointer_iterator
edge = sequence.back()->out_edges.begin();
edge != sequence.back()->out_edges.end(); ++edge)
{
if(unscheduled.count((*edge)->tail) != 0)
{
next = (*edge)->tail;
}
}
// rewind through fallthrough edges to find the beginning of the
// next chain of fall throughs
report(" restarting at " << next->label());
bool rewinding = true;
while (rewinding) {
rewinding = false;
for (const_edge_pointer_iterator
edge = next->in_edges.begin();
edge != next->in_edges.end(); ++edge) {
if ((*edge)->type == Edge::FallThrough) {
assertM(unscheduled.count((*edge)->head) != 0,
(*edge)->head->label()
<< " has multiple fallthrough branches.");
next = (*edge)->head;
report(" rewinding to " << next->label());
rewinding = true;
break;
}
}
}
sequence.push_back(next);
unscheduled.erase(next);
}
report(" added " << sequence.back()->label());
}
return sequence;
}
Shape *MakeLoop(BlockSet &Blocks, BlockSet& Entries, BlockSet &NextEntries) {
// Find the inner blocks in this loop. Proceed backwards from the entries until
// you reach a seen block, collecting as you go.
BlockSet InnerBlocks;
BlockSet Queue = Entries;
while (Queue.size() > 0) {
Block *Curr = *(Queue.begin());
Queue.erase(Queue.begin());
if (!contains(InnerBlocks, Curr)) {
// This element is new, mark it as inner and remove from outer
InnerBlocks.insert(Curr);
Blocks.erase(Curr);
// Add the elements prior to it
for (BlockSet::iterator iter = Curr->BranchesIn.begin(); iter != Curr->BranchesIn.end(); iter++) {
Queue.insert(*iter);
}
#if 0
// Add elements it leads to, if they are dead ends. There is no reason not to hoist dead ends
// into loops, as it can avoid multiple entries after the loop
for (BlockBranchMap::iterator iter = Curr->BranchesOut.begin(); iter != Curr->BranchesOut.end(); iter++) {
Block *Target = iter->first;
if (Target->BranchesIn.size() <= 1 && Target->BranchesOut.size() == 0) {
Queue.insert(Target);
}
}
#endif
}
}
assert(InnerBlocks.size() > 0);
for (BlockSet::iterator iter = InnerBlocks.begin(); iter != InnerBlocks.end(); iter++) {
Block *Curr = *iter;
for (BlockBranchMap::iterator iter = Curr->BranchesOut.begin(); iter != Curr->BranchesOut.end(); iter++) {
Block *Possible = iter->first;
if (!contains(InnerBlocks, Possible)) {
NextEntries.insert(Possible);
}
}
}
#if 0
// We can avoid multiple next entries by hoisting them into the loop.
if (NextEntries.size() > 1) {
BlockBlockSetMap IndependentGroups;
FindIndependentGroups(NextEntries, IndependentGroups, &InnerBlocks);
while (IndependentGroups.size() > 0 && NextEntries.size() > 1) {
Block *Min = NULL;
int MinSize = 0;
for (BlockBlockSetMap::iterator iter = IndependentGroups.begin(); iter != IndependentGroups.end(); iter++) {
Block *Entry = iter->first;
BlockSet &Blocks = iter->second;
if (!Min || Blocks.size() < MinSize) { // TODO: code size, not # of blocks
Min = Entry;
MinSize = Blocks.size();
}
}
// check how many new entries this would cause
BlockSet &Hoisted = IndependentGroups[Min];
bool abort = false;
for (BlockSet::iterator iter = Hoisted.begin(); iter != Hoisted.end() && !abort; iter++) {
Block *Curr = *iter;
for (BlockBranchMap::iterator iter = Curr->BranchesOut.begin(); iter != Curr->BranchesOut.end(); iter++) {
Block *Target = iter->first;
if (Hoisted.find(Target) == Hoisted.end() && NextEntries.find(Target) == NextEntries.end()) {
// abort this hoisting
abort = true;
break;
}
}
}
if (abort) {
IndependentGroups.erase(Min);
continue;
}
// hoist this entry
PrintDebug("hoisting %d into loop\n", Min->Id);
NextEntries.erase(Min);
for (BlockSet::iterator iter = Hoisted.begin(); iter != Hoisted.end(); iter++) {
Block *Curr = *iter;
InnerBlocks.insert(Curr);
Blocks.erase(Curr);
}
IndependentGroups.erase(Min);
}
}
#endif
PrintDebug("creating loop block:\n");
DebugDump(InnerBlocks, " inner blocks:");
DebugDump(Entries, " inner entries:");
DebugDump(Blocks, " outer blocks:");
DebugDump(NextEntries, " outer entries:");
LoopShape *Loop = new LoopShape();
Notice(Loop);
// Solipsize the loop, replacing with break/continue and marking branches as Processed (will not affect later calculations)
// A. Branches to the loop entries become a continue to this shape
for (BlockSet::iterator iter = Entries.begin(); iter != Entries.end(); iter++) {
Solipsize(*iter, Branch::Continue, Loop, InnerBlocks);
}
// B. Branches to outside the loop (a next entry) become breaks on this shape
for (BlockSet::iterator iter = NextEntries.begin(); iter != NextEntries.end(); iter++) {
Solipsize(*iter, Branch::Break, Loop, InnerBlocks);
}
// Finish up
Shape *Inner = Process(InnerBlocks, Entries, NULL);
Loop->Inner = Inner;
return Loop;
}