[PATCH] D23510: MachineScheduler: Restore TopCand after comparing top and bottom candidates

[Matthias Braun via llvm-commits]

MatzeB added a comment.

In https://reviews.llvm.org/D23510#516880, @tstellarAMD wrote:

> In https://reviews.llvm.org/D23510#515568, @MatzeB wrote:
>
> > Actually I am not so sure anymore. In cases where the topnode isn't picked the reason shouldn't matter, but we will reset the reason anyway next round when we compare against the top node again. Do you have an example that behaves differently with/without this patch? There may be something else wrong.
>
>
> pickNodeBidirectional() won't re-compare if topnode wasn't scheduled in the previous round.


It will always compare the top and bottom pick. Even though one of them is possibly saved from last round they will be subject to a comparison so outdates/old data in the reason field shouldn't affect the decision.


https://reviews.llvm.org/D23510