The parser spends the majority of its time allocating and freeing trees and stack nodes.
Also, the memory footprint of the AST is a significant concern when using tree-sitter
with large files. This library is already unlikely to work very well with source files
larger than 4GB, so representing rows, columns, byte lengths and child indices as
unsigned 32 bit integers seems like the right choice.
I think that state matching is the only correct strategy for incremental
node reuse that is compatible with the new error recovery algorithm. It's
also simpler than the sentential-form algorithm. With the compressed parse
tables, state matching shouldn't be too conservative of a test.
Before, any syntax error would cause the lexer to create an error
leaf node. This could happen even with a valid input, if the parse
stack had split and one particular version of the parse stack
failed to parse.
Now, an error leaf node is only created when the lexer cannot understand
part of the input stream at all. When a normal syntax error occurs,
the lexer just returns a token that is outside of the expected token
set, and the parser handles the unexpected token.
Suppose a parse state S has multiple actions for a terminal lookahead symbol A.
Then during incremental parsing, while in state S, the parser should not
reuse a non-terminal lookahead B where FIRST(B) contains A, because reusing B
might prematurely discard one of the possible actions that a batch parser
would have attempted in state S, upon seeing A as a lookahead.
Rather than letting the reduced tree become the new lookahead symbol,
and re-adding it to the stack via a subsequent shift action, just
add it to the stack as part of the reduce action. This is more in
line with the way LR is described traditionally.
