This makes it possible to run `make` or `cmake --build` in grammar
repos where `grammar.json` hasn't been included, and still get the parser
compilation to work.
**Problem:** `query.c` compares the current analysis state with the
previous analysis state to see if they are equal, so that it can return
early if so. This prevents redundant work. However, the comparison
function here differs from the one used for sorted insertion/lookup in
that it does not check any state data other than the child index. This
is problematic because it leads to infinite analysis when hidden nodes
have cycles.
**Solution:** Remove the custom comparison function, and apply the
insertion/lookup comparison function in place of it.
**NOTE:** This commit also changes the comparison function slightly, so
that some comparisons are reordered. Namely, for performance, it returns
early if the lhs depth is less than the rhs depth. Is this acceptable?
Tests still pass and nothing hangs in my testing, but it still seems
sketchy. Returning early if the lhs depth is greater than the rhs depth
does seem to make query analysis hang, weirdly enough... Keeping the
depth checks at the end of the loop also works, but it introduces a
noticeable performance regression (for queries that otherwise wouldn't
have had analysis cycles, of course).
This adds an `--evaluate-only` option to `tree-sitter generate`
so that it only does the evaluation of `grammar.js` to
`src/grammar.json`, without continuing on with the generation of
`src/parser.c` and related files.
It's a follow-up to #4580.
**Problem:** When encountering an invalid symbol at the beginning of the
file, the rust bindings attempt to index the character at position -1 of
the query source, which leads to an overflow and thus invalid character
index which causes a panic.
**Solution:** Bounds check the offset before performing the subtraction.
Rename libtree-sitter.lib target to libtree-sitter.dll.a which is
what MinGW actually uses, move the target after libtree-sitter.dll,
and make sure it is installed to the correct location.
Also avoid invoking `cc -dumpmachine` multiple times.
As the call to `symbol_is_used` does not depend
on the production, it is more efficient to call it
only once outside the loop over productions.
I'm not sure if `rustc` is able to do this optimization
on its own (it would need to know that the function
is pure, which sounds difficult in general).
Problem: After https://github.com/tree-sitter/tree-sitter/pull/4496, the
crater tests from nvim-treesitter no longer run on PRs with relevant changes.
Solution: Update the paths in the workflows to the new structure.
Also throttle the build step to 10 parallel tasks (reduces flakiness on
Windows).
Also raise the timeout limit to 30 minutes (it's more expensive to have
to restart flaky tests than to let them run).
