[llvm] [LoopInfo] Reuse DomTree's own DFS numbering in analyze() (PR #207650)
Alexis Engelke via llvm-commits
llvm-commits at lists.llvm.org
Mon Jul 6 00:49:19 PDT 2026
================
@@ -576,16 +576,23 @@ void PopulateLoopsDFS<BlockT, LoopT>::insertIntoLoop(BlockT *Block) {
/// insertions per block.
template <class BlockT, class LoopT>
void LoopInfoBase<BlockT, LoopT>::analyze(const DomTreeBase<BlockT> &DomTree) {
- // Postorder traversal of the dominator tree.
const DomTreeNodeBase<BlockT> *DomRoot = DomTree.getRootNode();
if constexpr (GraphHasNodeNumbers<const BlockT *>) {
ParentPtr = DomRoot->getBlock()->getParent();
BlockNumberEpoch = GraphTraits<ParentT>::getNumberEpoch(ParentPtr);
unsigned Max = GraphTraits<ParentT>::getMaxNumber(ParentPtr);
BBMap.resize(Max);
}
- for (auto DomNode : post_order(DomRoot)) {
+ // Visit dominator tree nodes in reverse preorder: like postorder, this
+ // guarantees a sub-loop is discovered before the outer loop.
+ DomTree.updateDFSNumbers();
+ SmallVector<const DomTreeNodeBase<BlockT> *, 32> PreorderNodes(
+ DomRoot->getDFSNumOut());
+ for (const DomTreeNodeBase<BlockT> *Node : DomTree.nodes())
+ PreorderNodes[Node->getDFSNumIn()] = Node;
----------------
aengelke wrote:
NB: this prevents skipping DFS number invalidation for eraseNode or intentionally keeping holes to avoid full renumbering on changes; this is the first use of these numbers that requires them to be consecutive. (I experimented with that before, eraseNode is not called often enough for this to make a substantial difference, often, the DT is recomputed afterwards anyway.) I don't particularly like this, but then, I don't particularly like LoopInfo and if it becomes a problem, it is easy to change later on.
https://github.com/llvm/llvm-project/pull/207650
More information about the llvm-commits
mailing list