This story begins in #8384, where we added a smart test for our syntax
highting, which run the algorithm on synthetic files of varying length
in order to guesstimate if the complexity is O(N^2) or O(N)-ish.
The test turned out to be pretty effective, and flagged #9031 as a
change that makes syntax highlighting accidentally quadratic. There was
much rejoicing, for the time being.
Then, lnicola asked an ominous question[1]: "Are we sure that the time
is linear right now?"
Of course it turned out that our sophisticated non-linearity detector
*was* broken, and that our syntax highlighting *was* quadratic.
Investigating that, many brave hearts dug deeper and deeper into the
guts of rust-analyzer, only to get lost in a maze of traits delegating
to traits delegating to macros.
Eventually, matklad managed to peel off all layers of abstraction one by
one, until almost nothing was left. In fact, the issue was discovered in
the very foundation of the rust-analyzer -- in the syntax trees.
Worse, it was not a new problem, but rather a well-know, well-understood
and event (almost) well-fixed (!) performance bug.
The problem lies within `SyntaxNodePtr` type -- a light-weight "address"
of a node in a syntax tree [3]. Such pointers are used by rust-analyzer all
other the place to record relationships between IR nodes and the
original syntax.
Internally, the pointer to a syntax node is represented by node's range.
To "dereference" the pointer, you traverse the syntax tree from the
root, looking for the node with the right range. The inner loop of this
search is finding a node's child whose range contains the specified
range. This inner loop was implemented by naive linear search over all
the children. For wide trees, dereferencing a single `SyntaxNodePtr` was
linear. The problem with wide trees though is that they contain a lot of
nodes! And dereferencing pointers to all the nodes is quadratic in the
size of the file!
The solution to this problem is to speed up the children search --
rather than doing a linear lookup, we can use binary search to locate
the child with the desired interval.
Doing this optimization was one of the motivations (or rather, side
effects) of #6857. That's why `rowan` grew the useful
`child_or_token_at_range` method which does exactly this binary search.
But looks like we've never actually switch to this method? Oups.
Lesson learned: do not leave broken windows in the fundamental infra.
Otherwise, you'll have to repeatedly re-investigate the issue, by
digging from the top of the Everest down to the foundation!
[1]: https://rust-lang.zulipchat.com/#narrow/stream/185405-t-compiler.2Frust-analyzer/topic/.60syntax_highlighting_not_quadratic.60.20failure/near/240811501
[2]: https://rust-lang.zulipchat.com/#narrow/stream/185405-t-compiler.2Frust-analyzer/topic/Syntax.20highlighting.20is.20quadratic
[3]: https://rust-lang.zulipchat.com/#narrow/stream/185405-t-compiler.2Frust-analyzer/topic/Syntax.20highlighting.20is.20quadratic/near/243412392
099b63e7c0