This is done by hoisting the instructions to handle ascending vs. non-ascending out of the inner loop in the bisection search. In our application, this function is the largest contributor to total runtime.
This change also fixes a latent bug potentially invoking undefined behavior (accessing element with index -1) if the table is empty.
This is done by hoisting the instructions to handle ascending vs. non-ascending out of the inner loop in the bisection search. In our application, this function is the largest contributor to total runtime.
This change also fixes a latent bug potentially invoking undefined behavior (accessing element with index -1) if the table is empty.
Pivoted flame graph before:
Pivoted flame graph after: