Examination


An estimation by JGNWG was confirmed.
It was that the Golliver's technique introduces a 2- or 4-fold penalty.
An observed performance decline was 1.53- 〜 2.98-fold.

We must make more experiments and analyze them.
The results seem to be inconsistent at a glance...

We need the guideline when implementing the `strictfp'.
Although several results are obtained, we have not gotten the guideline.
Do we have to analyze how machine (or μ-ops or RISC86) instructions are scheduled ?