08:18 ishitatsuyuki: can constant_folding run early in the optimization pipeline? an emulator developer contacted me about slow compilation of their shader (using spec constants) and it seems we're hitting the extremely heavy loop unroll pass before we const propagate.
08:20 ishitatsuyuki: or rather, const fold does run before loop unroll, but the order needs to be const fold -> eliminate dead CF -> loop unroll instead of dead CF -> const fold -> loop unroll
17:56 DemiMarie: ishitatsuyuki: If I recall correctly, constant folding, dead code elimination, and range inference should be combined into a single pass, because this generates better code than any combination of these passes.
17:59 DemiMarie: See https://factor-language.blogspot.com/2008/08/new-optimizer.html for why they should be combined.
18:16 konstantin: ishitatsuyuki: You could add a nir_opt_constant_folding call to spirv_to_nir that runs if spec constants were supplied. The pass if really cheap to run.
19:00 emersion: sounds like this could be somewhat related to DRM debug messages? https://lwn.net/Articles/984556/
19:22 mareko: robclark: the error message tells you where it's overridden
19:22 mareko: robclark: it's only missing the override keyword, and android is ignoring the -Wno option
19:23 mareko: it's also not even a warning on my machine
19:26 mareko: robclark: or disable -Werror for addrlib
19:44 robclark: mareko: I'd probably just add the `override` and then try to get the fix accepted "upstream".. but if that isn't an option then disabling `-Werror` makes sense
19:45 robclark: if it is code you can't fix then -Werror makes no sense
20:55 DemiMarie: Is there an estimate for when those who submitted proposals for XDC2024 will be notified if the proposal was accepted?