On Wed, Mar 19, 2025 at 2:06 AM Luis Gerhorst luis.gerhorst@fau.de wrote:
Thank you very much for having a look. Let me know whether the above resolves your concern.
In any case, should I separate patches 1-3 into another series?
Sorry for the delay. lsfmm was followed by the busy merge window.
Please rebase and repost the patches with the detailed explanation of how lfence works internally and it affects on the algorithm. I mistakenly thought that lfence is a load fence only. That it forces all prior loads to complete, but not the other insns. Since it's an absolute speculation barrier the algorithm appears sound. My only remaining reservation is a heuristic in this patch. If we don't do it, we wouldn't have to special case push_stack() too, right? Let's continue discussion in the new thread.