GitLab

all transitions and reductions on [error]. This leads to smaller
stars (if the grammar uses [error]), hence faster run time.

immediate feedback if it does not compile. Also fixed it to
properly show the compilation error message.

written for ocamlyacc and menhir.

size (6311, which means 8192) to a variable size, determined by the
size of the alphabet and the star size of the state. This seems to
lead to slightly slower behavior on small grammars and to slightly
faster running times and lower memory consumption on large grammars.

When emacs is open and reverts the buffer, this could help
avoid interference.

A copy of OCaml's hashtbl.ml, with micro-optimizations.
There is a tiny speed gain, but not really significant.

The speed gain is quite clear, but depends on the initial size of the hash
tables.

Will probably try other approaches soon.

Strangely, this slows things down and does not always save memory (!?).

(See patch in attic/word_and_lookahead.patch.)
This consistently saves 10% or 15% in space but costs just as much
in time, probably because hash-consing sees many more unique strings
(about 30x).

and performance of a new version with respect to the previous version.

This makes almost no difference: it saves 5 nodes out of 25K in
the stars, and saves 10K facts out of 7M facts (on ocaml.mly --lalr).