Il 11/05/2013 01:06, Anthony Liguori ha scritto: > The thing is, none of these casts should be for more than 1 level and > the patch I provided makes those casts almost free.
Assuming strings are collapsed (and that's why my original attempt to introduce the fast path only looked at the concrete classes: I was trying to optimize it within a device model). > I believe the reason it didn't fix the problem for Aurelien is because > the string addresses were different because of different compilation > units. I guess binutils doesn't collapse strings when linking. No, it doesn't. GCC can do that if you use link-time optimization, but binutils doesn't look into the data section. It is a black box. >>> As soon as we open up 1.6 for development, we face an immediate >>> regression in performance. So we need to fix the real problem here >>> anyway. >> >> No, we don't. We will simply have to live with a debugging vs. >> production tradeoff. > > I appreciate all of the arguments below. I'm very concerned about > reducing safety checks but am sympathetic to performance concerns. > > Here's what I would like to do. I'll apply 1-6 of your series which > gives us the infrastructure to disable qom casts. That at least let's > the code get tested in case we need it. Fair enough, thanks! Are you going to replace patch 3 with yours? (Though I don't think we need to check the parent, see above). > I will hold off applying patch 7 hoping that the patch I provided to > Aurelien solves the problem. If it doesn't and we're unable to find a > better solution, I'll apply patch 7 for the release. Let's please test virtio-net as well, it looks like Aurelien has the benchmark. /me goes swimming now. :) Paolo
