/// Sort the blocks, taking special care to make sure that loops are not
/// interrupted by blocks not dominated by their header.
/// TODO: There are many opportunities for improving the heuristics here.
/// Explore them.
static void SortBlocks(MachineFunction &MF, const MachineLoopInfo &MLI,
const MachineDominatorTree &MDT) {
// Prepare for a topological sort: Record the number of predecessors each
// block has, ignoring loop backedges.
MF.RenumberBlocks();
SmallVector<unsigned, 16> NumPredsLeft(MF.getNumBlockIDs(), 0);
for (MachineBasicBlock &MBB : MF) {
unsigned N = MBB.pred_size();
if (MachineLoop *L = MLI.getLoopFor(&MBB))
if (L->getHeader() == &MBB)
for (const MachineBasicBlock *Pred : MBB.predecessors())
if (L->contains(Pred))
--N;
NumPredsLeft[MBB.getNumber()] = N;
}
// Topological sort the CFG, with additional constraints:
// - Between a loop header and the last block in the loop, there can be
// no blocks not dominated by the loop header.
// - It's desirable to preserve the original block order when possible.
// We use two ready lists; Preferred and Ready. Preferred has recently
// processed sucessors, to help preserve block sequences from the original
// order. Ready has the remaining ready blocks.
PriorityQueue<MachineBasicBlock *, std::vector<MachineBasicBlock *>,
CompareBlockNumbers>
Preferred;
PriorityQueue<MachineBasicBlock *, std::vector<MachineBasicBlock *>,
CompareBlockNumbersBackwards>
Ready;
SmallVector<Entry, 4> Loops;
for (MachineBasicBlock *MBB = &MF.front();;) {
const MachineLoop *L = MLI.getLoopFor(MBB);
if (L) {
// If MBB is a loop header, add it to the active loop list. We can't put
// any blocks that it doesn't dominate until we see the end of the loop.
if (L->getHeader() == MBB)
Loops.push_back(Entry(L));
// For each active loop the block is in, decrement the count. If MBB is
// the last block in an active loop, take it off the list and pick up any
// blocks deferred because the header didn't dominate them.
for (Entry &E : Loops)
if (E.Loop->contains(MBB) && --E.NumBlocksLeft == 0)
for (auto DeferredBlock : E.Deferred)
Ready.push(DeferredBlock);
while (!Loops.empty() && Loops.back().NumBlocksLeft == 0)
Loops.pop_back();
}
// The main topological sort logic.
for (MachineBasicBlock *Succ : MBB->successors()) {
// Ignore backedges.
if (MachineLoop *SuccL = MLI.getLoopFor(Succ))
if (SuccL->getHeader() == Succ && SuccL->contains(MBB))
continue;
// Decrement the predecessor count. If it's now zero, it's ready.
if (--NumPredsLeft[Succ->getNumber()] == 0)
Preferred.push(Succ);
}
// Determine the block to follow MBB. First try to find a preferred block,
// to preserve the original block order when possible.
MachineBasicBlock *Next = nullptr;
while (!Preferred.empty()) {
Next = Preferred.top();
Preferred.pop();
// If X isn't dominated by the top active loop header, defer it until that
// loop is done.
if (!Loops.empty() &&
!MDT.dominates(Loops.back().Loop->getHeader(), Next)) {
Loops.back().Deferred.push_back(Next);
Next = nullptr;
continue;
}
// If Next was originally ordered before MBB, and it isn't because it was
// loop-rotated above the header, it's not preferred.
if (Next->getNumber() < MBB->getNumber() &&
(!L || !L->contains(Next) ||
L->getHeader()->getNumber() < Next->getNumber())) {
Ready.push(Next);
Next = nullptr;
continue;
}
break;
}
// If we didn't find a suitable block in the Preferred list, check the
// general Ready list.
if (!Next) {
// If there are no more blocks to process, we're done.
if (Ready.empty()) {
MaybeUpdateTerminator(MBB);
break;
}
for (;;) {
Next = Ready.top();
Ready.pop();
// If Next isn't dominated by the top active loop header, defer it until
// that loop is done.
if (!Loops.empty() &&
!MDT.dominates(Loops.back().Loop->getHeader(), Next)) {
Loops.back().Deferred.push_back(Next);
continue;
}
break;
}
}
// Move the next block into place and iterate.
Next->moveAfter(MBB);
MaybeUpdateTerminator(MBB);
MBB = Next;
}
assert(Loops.empty() && "Active loop list not finished");
MF.RenumberBlocks();
#ifndef NDEBUG
SmallSetVector<MachineLoop *, 8> OnStack;
// Insert a sentinel representing the degenerate loop that starts at the
// function entry block and includes the entire function as a "loop" that
// executes once.
OnStack.insert(nullptr);
for (auto &MBB : MF) {
assert(MBB.getNumber() >= 0 && "Renumbered blocks should be non-negative.");
MachineLoop *Loop = MLI.getLoopFor(&MBB);
if (Loop && &MBB == Loop->getHeader()) {
// Loop header. The loop predecessor should be sorted above, and the other
// predecessors should be backedges below.
for (auto Pred : MBB.predecessors())
assert(
(Pred->getNumber() < MBB.getNumber() || Loop->contains(Pred)) &&
"Loop header predecessors must be loop predecessors or backedges");
assert(OnStack.insert(Loop) && "Loops should be declared at most once.");
} else {
// Not a loop header. All predecessors should be sorted above.
for (auto Pred : MBB.predecessors())
assert(Pred->getNumber() < MBB.getNumber() &&
"Non-loop-header predecessors should be topologically sorted");
assert(OnStack.count(MLI.getLoopFor(&MBB)) &&
"Blocks must be nested in their loops");
}
while (OnStack.size() > 1 && &MBB == LoopBottom(OnStack.back()))
OnStack.pop_back();
}
assert(OnStack.pop_back_val() == nullptr &&
"The function entry block shouldn't actually be a loop header");
assert(OnStack.empty() &&
"Control flow stack pushes and pops should be balanced.");
#endif
}
/// Sort the blocks in RPO, taking special care to make sure that loops are
/// contiguous even in the case of split backedges.
///
/// TODO: Determine whether RPO is actually worthwhile, or whether we should
/// move to just a stable-topological-sort-based approach that would preserve
/// more of the original order.
static void SortBlocks(MachineFunction &MF, const MachineLoopInfo &MLI) {
// Note that we do our own RPO rather than using
// "llvm/ADT/PostOrderIterator.h" because we want control over the order that
// successors are visited in (see above). Also, we can sort the blocks in the
// MachineFunction as we go.
SmallPtrSet<MachineBasicBlock *, 16> Visited;
SmallVector<POStackEntry, 16> Stack;
MachineBasicBlock *EntryBlock = &*MF.begin();
Visited.insert(EntryBlock);
Stack.push_back(POStackEntry(EntryBlock, MF, MLI));
for (;;) {
POStackEntry &Entry = Stack.back();
SmallVectorImpl<MachineBasicBlock *> &Succs = Entry.Succs;
if (!Succs.empty()) {
MachineBasicBlock *Succ = Succs.pop_back_val();
if (Visited.insert(Succ).second)
Stack.push_back(POStackEntry(Succ, MF, MLI));
continue;
}
// Put the block in its position in the MachineFunction.
MachineBasicBlock &MBB = *Entry.MBB;
MBB.moveBefore(&*MF.begin());
// Branch instructions may utilize a fallthrough, so update them if a
// fallthrough has been added or removed.
if (!MBB.empty() && MBB.back().isTerminator() && !MBB.back().isBranch() &&
!MBB.back().isBarrier())
report_fatal_error(
"Non-branch terminator with fallthrough cannot yet be rewritten");
if (MBB.empty() || !MBB.back().isTerminator() || MBB.back().isBranch())
MBB.updateTerminator();
Stack.pop_back();
if (Stack.empty())
break;
}
// Now that we've sorted the blocks in RPO, renumber them.
MF.RenumberBlocks();
#ifndef NDEBUG
SmallSetVector<MachineLoop *, 8> OnStack;
// Insert a sentinel representing the degenerate loop that starts at the
// function entry block and includes the entire function as a "loop" that
// executes once.
OnStack.insert(nullptr);
for (auto &MBB : MF) {
assert(MBB.getNumber() >= 0 && "Renumbered blocks should be non-negative.");
MachineLoop *Loop = MLI.getLoopFor(&MBB);
if (Loop && &MBB == Loop->getHeader()) {
// Loop header. The loop predecessor should be sorted above, and the other
// predecessors should be backedges below.
for (auto Pred : MBB.predecessors())
assert(
(Pred->getNumber() < MBB.getNumber() || Loop->contains(Pred)) &&
"Loop header predecessors must be loop predecessors or backedges");
assert(OnStack.insert(Loop) && "Loops should be declared at most once.");
} else {
// Not a loop header. All predecessors should be sorted above.
for (auto Pred : MBB.predecessors())
assert(Pred->getNumber() < MBB.getNumber() &&
"Non-loop-header predecessors should be topologically sorted");
assert(OnStack.count(MLI.getLoopFor(&MBB)) &&
"Blocks must be nested in their loops");
}
while (OnStack.size() > 1 && &MBB == LoopBottom(OnStack.back()))
OnStack.pop_back();
}
assert(OnStack.pop_back_val() == nullptr &&
"The function entry block shouldn't actually be a loop header");
assert(OnStack.empty() &&
"Control flow stack pushes and pops should be balanced.");
#endif
}
