| 5 Mar 2026 |
 hexa | fair | 15:21:41 |
 emily | would be nice if we had deterministic PGO | 15:22:47 |
| emily | but it's hard | 15:22:50 |
 Lun | is hand tuned small top % wins reviewer legible proftext viable or would leaving most things with no instrumentation make it pessimize a lot. | 15:28:14 |
| emily | no idea :) | 15:30:09 |
| emily | I do know there are concerns about compilers not being hardened against malicious profile data (so they're essentially load-bearing trusted blobs even if we made them fixed), but I assume making it human-auditable would address that concern | 15:30:55 |
| Lun | will have to test it some time and not while having breakfast in a hotel lobby ig | 15:31:38 |
|  Caturday Reed changed their display name from Steve Reed to Caturday Reed. | 15:35:28 |
| Caturday Reed changed their profile picture. | 15:35:32 |
 vcunat | No, I don't know. But I'm quite sure that removing x86_64-darwin will help extremely. | 15:36:34 |
| emily | yeah for sure. just wondering what else might have gone on to make things worse recently | 15:37:08 |
| emily | we did just land the eval warning for x86_64-darwin | 15:37:14 |
| emily | need to get back to my stack that rips it out for 26.05 soon | 15:37:25 |
| vcunat | Maybe the last couple of cycles were slower for darwin, but it seems so now even for staging-next-25.11 and this kind of changes weren't backported, right? | 15:37:46 |
| vcunat | * Maybe the last couple of cycles felt slower for darwin, but it seems so now even for staging-next-25.11 and this kind of changes weren't backported, right? | 15:37:56 |
| emily | the change hit 25.11 | 15:38:10 |
| emily | so it was already present at release, AIUI | 15:38:20 |
| vcunat | Oh, OK. | 15:38:20 |
| emily | so if 25.11 got slower and wasn't always slow then it's probably not that | 15:38:33 |
| emily | but if 25.11 was always slower than 25.05 it could be that | 15:38:39 |
| emily | (unless it's "got slower once the builders were updated to 25.11" 😅) | 15:39:23 |
| vcunat | October was quite a long ago. | 15:39:31 |
| emily | last couple cycles sounds more like OS upgrade to me, though I forget when the last one of those was. | 15:39:42 |
| vcunat | I mean, if it's a nixpkgs change, you could pick a random package and measure the time a year ago and now. | 15:40:30 |
| vcunat | If it's a well measurable difference, then even bisection is viable. | 15:40:50 |
| vcunat | (--first-parent bisection might be relatively low-effort) | 15:41:38 |
| vcunat | (until you hit a staging* merge - and then it's harder to dig inside) | 15:42:18 |
| emily | right. I probably shouldn't get sidetracked investigating right now, was just wondering if there was anything obvious I could correlate with when things started to get slower | 15:42:19 |
| vcunat | * (until you hit the staging* merge containing the regression - and then it's harder to dig inside) | 15:42:34 |
| vcunat | No, I don't think we have that. | 15:42:50 |
