// Check that 'BB' doesn't have any uses outside of the 'L'
static bool isBlockInLCSSAForm(const Loop &L, const BasicBlock &BB,
DominatorTree &DT) {
for (const Instruction &I : BB) {
// Tokens can't be used in PHI nodes and live-out tokens prevent loop
// optimizations, so for the purposes of considered LCSSA form, we
// can ignore them.
if (I.getType()->isTokenTy())
continue;
for (const Use &U : I.uses()) {
const Instruction *UI = cast<Instruction>(U.getUser());
const BasicBlock *UserBB = UI->getParent();
if (const PHINode *P = dyn_cast<PHINode>(UI))
UserBB = P->getIncomingBlock(U);
// Check the current block, as a fast-path, before checking whether
// the use is anywhere in the loop. Most values are used in the same
// block they are defined in. Also, blocks not reachable from the
// entry are special; uses in them don't need to go through PHIs.
if (UserBB != &BB && !L.contains(UserBB) &&
DT.isReachableFromEntry(UserBB))
return false;
}
}
return true;
}
/// isLCSSAForm - Return true if the Loop is in LCSSA form
bool Loop::isLCSSAForm(DominatorTree &DT) const {
// Sort the blocks vector so that we can use binary search to do quick
// lookups.
SmallPtrSet<BasicBlock*, 16> LoopBBs(block_begin(), block_end());
for (block_iterator BI = block_begin(), E = block_end(); BI != E; ++BI) {
BasicBlock *BB = *BI;
for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;++I)
for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;
++UI) {
User *U = *UI;
BasicBlock *UserBB = cast<Instruction>(U)->getParent();
if (PHINode *P = dyn_cast<PHINode>(U))
UserBB = P->getIncomingBlock(UI);
// Check the current block, as a fast-path, before checking whether
// the use is anywhere in the loop. Most values are used in the same
// block they are defined in. Also, blocks not reachable from the
// entry are special; uses in them don't need to go through PHIs.
if (UserBB != BB &&
!LoopBBs.count(UserBB) &&
DT.isReachableFromEntry(UserBB))
return false;
}
}
return true;
}
static bool ProcessBlock(BasicBlock &BB, DominatorTree &DT, LoopInfo &LI,
AAResults &AA) {
// Can't sink anything out of a block that has less than two successors.
if (BB.getTerminator()->getNumSuccessors() <= 1) return false;
// Don't bother sinking code out of unreachable blocks. In addition to being
// unprofitable, it can also lead to infinite looping, because in an
// unreachable loop there may be nowhere to stop.
if (!DT.isReachableFromEntry(&BB)) return false;
bool MadeChange = false;
// Walk the basic block bottom-up. Remember if we saw a store.
BasicBlock::iterator I = BB.end();
--I;
bool ProcessedBegin = false;
SmallPtrSet<Instruction *, 8> Stores;
do {
Instruction *Inst = &*I; // The instruction to sink.
// Predecrement I (if it's not begin) so that it isn't invalidated by
// sinking.
ProcessedBegin = I == BB.begin();
if (!ProcessedBegin)
--I;
if (isa<DbgInfoIntrinsic>(Inst))
continue;
if (SinkInstruction(Inst, Stores, DT, LI, AA)) {
++NumSunk;
MadeChange = true;
}
// If we just processed the first instruction in the block, we're done.
} while (!ProcessedBegin);
return MadeChange;
}
/// isLCSSAForm - Return true if the Loop is in LCSSA form
bool Loop::isLCSSAForm(DominatorTree &DT) const {
for (block_iterator BI = block_begin(), E = block_end(); BI != E; ++BI) {
BasicBlock *BB = *BI;
for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;++I)
for (Use &U : I->uses()) {
Instruction *UI = cast<Instruction>(U.getUser());
BasicBlock *UserBB = UI->getParent();
if (PHINode *P = dyn_cast<PHINode>(UI))
UserBB = P->getIncomingBlock(U);
// Check the current block, as a fast-path, before checking whether
// the use is anywhere in the loop. Most values are used in the same
// block they are defined in. Also, blocks not reachable from the
// entry are special; uses in them don't need to go through PHIs.
if (UserBB != BB &&
!contains(UserBB) &&
DT.isReachableFromEntry(UserBB))
return false;
}
}
return true;
}
/// Given \p BBs as input, find another set of BBs which collectively
/// dominates \p BBs and have the minimal sum of frequencies. Return the BB
/// set found in \p BBs.
static void findBestInsertionSet(DominatorTree &DT, BlockFrequencyInfo &BFI,
BasicBlock *Entry,
SmallPtrSet<BasicBlock *, 8> &BBs) {
assert(!BBs.count(Entry) && "Assume Entry is not in BBs");
// Nodes on the current path to the root.
SmallPtrSet<BasicBlock *, 8> Path;
// Candidates includes any block 'BB' in set 'BBs' that is not strictly
// dominated by any other blocks in set 'BBs', and all nodes in the path
// in the dominator tree from Entry to 'BB'.
SmallPtrSet<BasicBlock *, 16> Candidates;
for (auto BB : BBs) {
// Ignore unreachable basic blocks.
if (!DT.isReachableFromEntry(BB))
continue;
Path.clear();
// Walk up the dominator tree until Entry or another BB in BBs
// is reached. Insert the nodes on the way to the Path.
BasicBlock *Node = BB;
// The "Path" is a candidate path to be added into Candidates set.
bool isCandidate = false;
do {
Path.insert(Node);
if (Node == Entry || Candidates.count(Node)) {
isCandidate = true;
break;
}
assert(DT.getNode(Node)->getIDom() &&
"Entry doens't dominate current Node");
Node = DT.getNode(Node)->getIDom()->getBlock();
} while (!BBs.count(Node));
// If isCandidate is false, Node is another Block in BBs dominating
// current 'BB'. Drop the nodes on the Path.
if (!isCandidate)
continue;
// Add nodes on the Path into Candidates.
Candidates.insert(Path.begin(), Path.end());
}
// Sort the nodes in Candidates in top-down order and save the nodes
// in Orders.
unsigned Idx = 0;
SmallVector<BasicBlock *, 16> Orders;
Orders.push_back(Entry);
while (Idx != Orders.size()) {
BasicBlock *Node = Orders[Idx++];
for (auto ChildDomNode : DT.getNode(Node)->getChildren()) {
if (Candidates.count(ChildDomNode->getBlock()))
Orders.push_back(ChildDomNode->getBlock());
}
}
// Visit Orders in bottom-up order.
using InsertPtsCostPair =
std::pair<SmallPtrSet<BasicBlock *, 16>, BlockFrequency>;
// InsertPtsMap is a map from a BB to the best insertion points for the
// subtree of BB (subtree not including the BB itself).
DenseMap<BasicBlock *, InsertPtsCostPair> InsertPtsMap;
InsertPtsMap.reserve(Orders.size() + 1);
for (auto RIt = Orders.rbegin(); RIt != Orders.rend(); RIt++) {
BasicBlock *Node = *RIt;
bool NodeInBBs = BBs.count(Node);
SmallPtrSet<BasicBlock *, 16> &InsertPts = InsertPtsMap[Node].first;
BlockFrequency &InsertPtsFreq = InsertPtsMap[Node].second;
// Return the optimal insert points in BBs.
if (Node == Entry) {
BBs.clear();
if (InsertPtsFreq > BFI.getBlockFreq(Node) ||
(InsertPtsFreq == BFI.getBlockFreq(Node) && InsertPts.size() > 1))
BBs.insert(Entry);
else
BBs.insert(InsertPts.begin(), InsertPts.end());
break;
}
BasicBlock *Parent = DT.getNode(Node)->getIDom()->getBlock();
// Initially, ParentInsertPts is empty and ParentPtsFreq is 0. Every child
// will update its parent's ParentInsertPts and ParentPtsFreq.
SmallPtrSet<BasicBlock *, 16> &ParentInsertPts = InsertPtsMap[Parent].first;
BlockFrequency &ParentPtsFreq = InsertPtsMap[Parent].second;
// Choose to insert in Node or in subtree of Node.
// Don't hoist to EHPad because we may not find a proper place to insert
// in EHPad.
// If the total frequency of InsertPts is the same as the frequency of the
// target Node, and InsertPts contains more than one nodes, choose hoisting
// to reduce code size.
if (NodeInBBs ||
(!Node->isEHPad() &&
(InsertPtsFreq > BFI.getBlockFreq(Node) ||
(InsertPtsFreq == BFI.getBlockFreq(Node) && InsertPts.size() > 1)))) {
ParentInsertPts.insert(Node);
ParentPtsFreq += BFI.getBlockFreq(Node);
} else {
ParentInsertPts.insert(InsertPts.begin(), InsertPts.end());
ParentPtsFreq += InsertPtsFreq;
}
}
}
