01:54 airlied: wow llvmpipe on llvm22 was quite buste
01:54 airlied: busted
02:44 HdkR: Consistently broken api woo
03:54 llyyr: rusticl fails to build on latest rustc/bindgen with a bunch of error[E0080]: attempt to compute `1_usize - 8_usize`, which would overflow in rusticl_opencl_bindings
07:05 tursulin: shivamklr: entity->rq? I did not plan to work on that so feel free to study it. I am not sure there is much to do apart maybe annotating some places to make it clearer why entity->rq is guaranteed to be stable at a specific dereference site.
09:52 karolherbst: llyyr: yeah.. is it inside the test code bindgen generates?
09:53 karolherbst: bindgen sometimes has a hard time with align attributes
09:53 karolherbst: and e.g. the tests for the vector types were almost always broken
09:53 llyyr: I thought it was only rusticl_opencl_bindings.rs, but the generated rusticl_mesa_bindings.rs is also affected
09:54 llyyr: seems to be related to libclang update to 22
09:54 llyyr: haven't looked into it further
09:55 karolherbst: mhhh hopefully it's something silly we can ignore
10:02 karolherbst: llyyr: what rustc and bindgen version are you on?
10:03 llyyr: rustc 1.94.1 (e408947bf 2026-03-25) and bindgen 0.72.0
10:03 karolherbst: airlied: have you seen this issue? Not sure if that's an llvmpipe bug or not.. but seems to be llvm22 related https://gitlab.freedesktop.org/mesa/mesa/-/work_items/15355
10:11 karolherbst: yeah.. looks like this is llvm22 related indeed
10:12 karolherbst: llyyr: have you checked if 0.72.1 fixes it? Apparently there were some llvm22 fixes
10:12 karolherbst: for bindgen I mean
10:13 llyyr: I can try, my distro haven't shipped 0.72.1 yet
10:15 karolherbst: I found this issue: https://github.com/rust-lang/rust-bindgen/issues/3264
10:16 karolherbst: and apparently somebody already reported it: https://gitlab.freedesktop.org/mesa/mesa/-/work_items/13757
10:25 llyyr: karolherbst: 0.72.1 does fix it, thanks!
10:25 llyyr: probably should raise the minimum in mesa
11:10 shivamklr: tursulin Thanks, that’s useful. I’ll audit the `entity->rq` dereference sites first and check whether the stability assumptions are already guaranteed. If it turns out to mostly need clarification rather than locking, I’ll shape it that way and send an RFC once I have something concrete.
11:37 shivamklr: tursulin: (sorry, forgot the colon)
13:33 karolherbst: llyyr: we have a few checks on the meson level to detect broken bindgen versions, could add another one I guess...
13:33 karolherbst: or yeah.. just require 0.72.1 it's already old
13:33 karolherbst: like almost a year
14:13 karolherbst: didn't we had an optimization where the first iteration of a loop gets extracted or something?
14:15 pendingchaos: opt_loop_peel_initial_break()?
14:24 llyyr: I realized my distro already had shipped 0.72.1, it's just that I had another bindgen in my ~/.cargo/bin which was shadowing the /usr/bin one
14:24 llyyr: self inflicted issue I suppose
14:32 karolherbst: ahh yeah, opt_loop_peel_initial_break. I should play around with that one
15:10 karolherbst: something I just saw: %18 = iadd %17 (0x80), %9; div 1 %20 = ult %18, %19 (0x10). I think that ult is always false :)
15:14 karolherbst: ehh wait.. it could overflow mhh, but it's initially 0...