Push performance much closer to the theoretical optimum.
Issue shall be closed when all the obvious optimizations have been applied.
The decode algorithm should be revisited. Does it copy too much information around that could be shared, and would sharing subtrees result in a performance gain?
Also, it should theoretically be possible to parse the Dyck_1 grammar without backtracking, no? But the current algorithm can't do it. However, the parsing algorithm can easily be replaced without touching the rest of the decoder.
Push performance much closer to the theoretical optimum.
Issue shall be closed when all the obvious optimizations have been applied.
The decode algorithm should be revisited. Does it copy too much information around that could be shared, and would sharing subtrees result in a performance gain?
Also, it should theoretically be possible to parse the Dyck_1 grammar without backtracking, no? But the current algorithm can't do it. However, the parsing algorithm can easily be replaced without touching the rest of the decoder.
Push performance much closer to the theoretical optimum.
Issue shall be closed when all the obvious optimizations have been applied.
The decode algorithm should be revisited. Does it copy too much information around that could be shared, and would sharing subtrees result in a performance gain?
Also, it should theoretically be possible to parse the Dyck_1 grammar without backtracking, no? But the current algorithm can't do it. However, the parsing algorithm can easily be replaced without touching the rest of the decoder.
Good news: specialization to binary fractions worked.
As for the last point, Greibach normal form could help.