Address `ts_node_parent` performance issues
andreaphylum opened this issue · comments
When traversing a tree upwards, we're currently using the ts_node_parent function rather than the Cursor API, which is inefficient as it traverses the whole tree at every invocation; this leads to performance issues that are apparent even on relatively small test fixtures.
We should investigate whether using the Cursor API provides significant advantages, though it is likely that a single traversal will be necessary as there is no API for reaching a given node in a cursor in O(1) (see this).
There is no apparent efficient way of traversing a syntax tree upwards, as tree-sitter has some crucial functionality missing from its Cursor API (e.g. the possibility of retrieving a nearby node without moving the cursor).
An alternative, more efficient approach would be to visit the tree downwards, annotating nodes and caching along the way; but that requires rethinking the algorithms and giving up an important layer of abstraction, as referring to nodes' parents is very useful in understanding the code.
I suggest we postpone pursuing this optimization in the interest of feature progress, but I think it is important to eventually get to it.