20:19 <anarsoul> enunes: for your "remove nir_legacy" MR, can you share examples of the shaders that regressed?

20:24 <enunes> from what I saw these are the shaders that arrived to the backend not with nir ops to be folded but with the nir flags like "legacy_fneg" set and now that we changed the trivialize pass don't have that anymore

20:24 <enunes> so now they are the cases like neg on register which are hard to optimize after we convert everything to ppir ir

20:25 <enunes> lets see if I can find a relatively simple one

20:26 <anarsoul> it's 12.5% increase in instructions for worst case

20:28 <anarsoul> enunes: are we getting hit again by a non-existing copy prop pass for ppir? :)

20:28 <enunes> I guess that is like 7->8 or something, not 100->125

20:29 <anarsoul> enunes: in absolute numbers worst hit is 15 instructions

20:29 <anarsoul> HURT stats (abs) min: 1 max: 15 x̄: 3.57 x̃: 2

20:29 <enunes> definitely feels limited by ppir ir representation indeed

20:31 <enunes> this is the full report of the stats I posted there https://paste.centos.org/view/raw/6d8983ed

20:33 <anarsoul> instructions HURT: shaders/godot3.4/31-18.shader_test MESA_SHADER_FRAGMENT: 231 -> 246 (6.49%) (spills: 7 -> 15 (114.29%); fills: 24 -> 38 (58.33%))instructions HURT: shaders/yofrankie/105.shader_test MESA_SHADER_FRAGMENT: 159 -> 172 (8.18%)

20:34 <enunes> yeah I guess this just snowballs on a small change and spills become worse

20:34 <anarsoul> anyway, your changes are straightforward and clean, so I'm not oppositing to merge it

20:35 <enunes> with the final change it is true that some of the new foldings mask more of the losses of the new pass vs nir folding

20:57 <enunes> I guess while I'm at it I'll take another look to see if we can still save some of the register cases, today I was just focusing on tracking down the remaining regressions with that current code