[PATCH] D65506: [MachineScheduler] improve reuse of 'releaseNode'method

Florian Hahn via Phabricator via llvm-commits llvm-commits at lists.llvm.org
Fri Nov 22 08:11:12 PST 2019


fhahn added a comment.

In D65506#1719962 <https://reviews.llvm.org/D65506#1719962>, @DoktorC wrote:

> Hi @fhahn ,
>  sorry for the long silence period.
>
> I tracked down the problem: it is how the `Pending` queue is explored with this patch (from the end).
>  Clearly, this affects which nodes move first from pending to available state; this is critical when
>  the `Available` queue reaches the maximum size.
>
> I can re-write the queue exploration to match the pre-patch behaviour.
>  However, I'm wondering if it is a sign of "bad" design: shouldn't the pending queue explore the best
>  candidates to move in the available queue?


I think it would be best keep the old behavior in this patch and potentially fix the iteration order in a follow up. Changes to the scheduler can lead to subtle regressions, so it's best to keep the changes isolated.


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D65506/new/

https://reviews.llvm.org/D65506





More information about the llvm-commits mailing list