The memory allocated by 'closure' (and some data such as 'fderives')
is used to computed a state's full itemset from its core. This is
needed during the construction of the LR(0) automaton, and the memory
is reclaimed immediately afterwards.
Unfortunately the reports (graph, text, xml) also need this
information when describing the states with their full itemsets. As a
consequence the memory was allocated again, fderives computed again
too, and more --trace reports are generated which only duplicate what
was already reported.
Stop that. It does mean that we release the memory later (hence the
peak memory usage is higher now), but I don't think that's a problem
today.
* src/lr0.c (generate_states): Don't call closure_free.
* src/state.c (states_free): Do it here.
(for symmetry with closure_new which is called in generate_states).
* src/print-graph.c, src/print-xml.c, src/print.c: You can now expect
the closure module to be functional.
This is more consistent with the other files.
* closure.h, closure.c (new_closure, free_closure): Rename as...
(closure_new, closure_free): this.
Adjust dependencies.
This will make it easier to add new elements (that might already be
part of shift_symbol) without having to worry about the size of
shift_symbol (which is currently a fixed size vector).
I could not measure any significant differences in performances in the
generation of LR(0) automaton (benched on gramamrs of Ruby, C, and C++).
* src/lr0.c (shift_symbol): Make it a bitset.
