| 10 Apr 2026 |
 Atemu | tbh. if something works on one non-4k, it probably also works on any page size but a decent amount probably only works on 4k | 09:28:20 |
 Yureka (she/her) | I would really consider this a bug | 09:29:14 |
 K900 | I think this is a bad idea fwiw | 09:29:27 |
| K900 | Huge Hydra workload for what | 09:29:36 |
| K900 | Buggy software | 09:29:38 |
| K900 | That is mostly near trivial to fix | 09:29:47 |
 vcunat | I don't think it's our choice anyway. | 09:29:47 |
| vcunat | It's for parts like gcc/libc what they consider too different. | 09:30:13 |
| Atemu | I mean, it doesn't mean we need to have a separate binary cache for it | 09:30:36 |
| Yureka (she/her) | what would be different about it then? | 09:30:54 |
| Atemu | Hm, good point. It'd only matter in such broken packages where you'd be able to respond accordingly and there's probably packages where you can set page size at compile time | 09:32:32 |
| Yureka (she/her) | Also please keep in mind that this is just a wild guess that openblas 0.3.32 produces garbage on non-4k pagesize systems. My sample size is 1 4k pagesize system and 2 non-4k pagesize systems, so not a very big sample size. | 09:32:45 |
| Yureka (she/her) | It might also be any instruction set extension or other difference between those processors which happens to correlate with 4k/non-4k | 09:33:27 |
| Yureka (she/her) | It's just that indeed a lot of software is buggy on non-4k pagesize systems, but this doesn't really change anything about how we should build it | 09:34:12 |
| Yureka (she/her) | Btw, this is not even really a property of the hardware. If I understand it correctly I can also run my Macbook cpu on a 4k pagesize kernel if I don't care about graphics or display. | 09:35:41 |
| Yureka (she/her) | And you can decide to run any/most aarch64 hosts on 16k or 64k if you'd like to | 09:36:10 |
| K900 | Yeah aarch64 baseline includes support for bigger page sizes | 09:36:36 |
| K900 | It's just not particularly useful in most cases | 09:36:46 |
| K900 | Except Asahi machines because Apple IOMMU weirdness | 09:36:54 |
| Yureka (she/her) | Well depends on your workload | 09:36:58 |
| Yureka (she/her) | and tradeoffs | 09:37:14 |
| vcunat | More precisely, POSIX chose that page size is not a compile-time constant. (I believe, after a bit of looking around) | 09:37:59 |
| vcunat | * More precisely, POSIX chose that page size is not a compile-time constant. (I believe, after a bit of looking around) I think it would be too painful to diverge. | 09:38:31 |
| vcunat | * More precisely, POSIX chose that page size is not a compile-time constant. (I believe, after a bit of looking around) I think it would be too painful to differ on this point. | 09:38:46 |
| Yureka (she/her) | But yes it's definitely openblas 0.3.31 -> 0.3.32. Reverting that, I can also build scipy on my macbook again. I'll bisect openblas next. | 09:49:28 |
| Yureka (she/her) | 3f6e928d34aca977bd5d4191e6d2c2338a342db5 | 10:35:01 |
| Yureka (she/her) | Declare result as volatile to keep compilers from optimizing it out | 10:35:06 |
| Yureka (she/her) | Redacted or Malformed Event | 10:35:10 |
| Yureka (she/her) | Redacted or Malformed Event | 10:35:25 |
| Yureka (she/her) | https://github.com/OpenMathLib/OpenBLAS/issues/5708 | 10:36:10 |
