Hi Kyrill > You first need to check that targetm.sched.issue_rate is non-null. > I wonder if the schedule issue_rate is too coarse. The midend also uses > reassociation width as an estimate of CPU parallelism. On aarch64 at least we > describe different widths for int, fp, and vector reassociation. > Have you considered using that hook instead?
Thanks for the suggestion. I have updated the draft to use targetm.sched.reassociation_width for integer, floating-point, and vector operations separately. I retained issue_rate, with a null check, only as a global per-level limit, since reassociation width may exceed the actual issue width due to latency hiding. Regards, Wang Jue
