08:24 notthatclippy[d]: _lyude[d]: I think all the first stage ucodes have to be contiguous. Second stage stuff, like gsp.bin can be discontig.
08:26 notthatclippy[d]: We could probably change the generic bootloader to support radix3, but I imagine just reserving a small chunk of contig space is easier. You just have to have one chunk for the max ucode size and swap out what resides there as different things get booted
16:17 karolherbst[d]: oof.. I found a case where my RA Phi Vec thing reduces instructions from 2535 to 2159, it's all mov instructions 🥲
16:19 karolherbst[d]: mhenning[d]: did you had a better idea for my pass or something? I think I've chatted with somebody about it, but can't really remember
16:20 mhenning[d]: Yeah, I think we talked about it but I forget what we said
16:21 karolherbst[d]: anyway.. the issue is kinda that the register get allocated before the vec use is seen
16:22 karolherbst[d]: but also not sure if we've talked before or after my latest rework of the opt
16:23 mhenning[d]: yeah, we need to do some analysis ahead of time so we have that info at allocation time
16:23 karolherbst[d]: yeah, that's what my code is doing
16:23 karolherbst[d]: it's just... ugly I think 😄
16:23 karolherbst[d]: but it does integrate into the phi webs now
16:23 karolherbst[d]: when allocating
16:23 mhenning[d]: I can take a look at the MR if you want.
16:24 karolherbst[d]: yeah, maybe for the best
16:24 karolherbst[d]: the part I don't like is the stuff inside `PhiWebs::new`
16:24 karolherbst[d]: everything else is fine I think
16:32 karolherbst[d]: and I think https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/37998 is ready, just need a final review.
16:48 _lyude[d]: notthatclippy[d]: Yeah - that's what I'm doing now! It seems to work, though I'll know for sure after my workday today since I'm currently testing such a fix