| 30 Jun 2026 |
 Grimmauld (any/all) | Can't we override the llvm going into rustc as a hack? | 17:32:23 |
 Lach | But it seems to be best solution given that rustc expects patches that were not merged in upstream llvm at the time | 17:32:35 |
| Lach | I'm not sure if there is more and this one commit is enough | 17:33:08 |
 whispers [& it/fae] | it's 40000 linux/30000 darwin roughly. pretty big. (https://github.com/NixOS/nixpkgs/pull/537054) | 17:36:21 |
| Lach | Actually, this fix is in llvm 22
We can switch rustc to build with llvmPackages_22 maybe? | 17:38:13 |
 emily | do any of the major cryptography libraries use AVX512 runtime detection? | 17:38:19 |
| emily | imo if yes → eat the rebuilds, if no → package 1.96.1 separately for this cycle and override stuff we know will break to use it | 17:38:37 |
| Lach | I think it affects only _mm512_dpbusd_epi32 intrinsic specifically | 17:39:05 |
| emily | looks like zlib-rs uses it. ouch. | 17:39:38 |
| Lach | zlib_rs, ndarray are using it | 17:39:55 |
| emily | https://github.com/trifectatechfoundation/zlib-rs/blob/5a96dcf8f36644074ca604dadae36591de5551a5/zlib-rs/src/adler32/avx512.rs#L42 or not? "emulated"… | 17:39:58 |
| emily | if it actually affects zlib-rs then that's serious enough we should probably reroll Rust | 17:40:35 |
| Lach | https://github.com/AdaWorldAPI/ndarray/blob/8c381a61040884180f3909d755f45d129d14cd4c/src/hpc/int8_tile_gemm.rs#L152 | 17:40:38 |
| emily | it seems like it only calls this "emulated" version though | 17:41:05 |
| Lach | https://github.com/trifectatechfoundation/zlib-rs/blob/5a96dcf8f36644074ca604dadae36591de5551a5/zlib-rs/src/adler32/avx512_vnni.rs#L84 | 17:42:12 |
| Lach | vnni path uses the real one | 17:42:49 |
| emily | aha, ok. separate files. | 17:42:48 |
| emily | can we get Linux to fake CPUID? 🤪 | 17:43:05 |
| Lach | It won't help since it breaks at compile time | 17:43:24 |
| Lach | llvm 21 doesn't understand the code that rustc 1.96 emits | 17:43:38 |
| Lach | It needs this patch https://github.com/rust-lang/llvm-project/commit/94e2c19f86a699d7a19ff0f4130b696699189c8d | 17:43:55 |
| emily | oh, it's just a build failure? | 17:44:15 |
| Lach | Yes | 17:44:23 |
| emily | that's totally fine then. we can probably fix it for packages that break | 17:44:26 |
 K900 | Two rustcs? | 17:44:35 |
| K900 | That's going to suuuuuuuck | 17:44:40 |
| K900 | And there's the 1.96.1 miscompilation | 17:44:47 |
| Lach | I don't think so, as with ndarray and other packages this issue would be deep into the dependency graph | 17:44:48 |
| emily | I guess if it's everything using zlib-rs, yeah… | 17:44:50 |
| K900 | Which is much more subtle | 17:44:49 |
