| 20 May 2025 |
 emily | I think we can disable tests on old versions of those. | 01:40:43 |
 hexa | this could be all that's needed | 01:42:57 |
| hexa | do I start contributing to llvm now? 😫 | 01:43:36 |
| hexa | also tblgen iirc, but i'd have to check | 01:44:14 |
| emily | they don't maintain the old versions at all | 01:43:49 |
| hexa | hm no | 01:45:31 |
| emily | everything but 20 is technically EOL | 01:43:57 |
| hexa | ah, great | 01:47:13 |
| emily | which is also why as many of them as possible should go away in 25.11 | 01:44:12 |
| emily | so some of those Python packages may be casualties anyway | 01:44:22 |
| hexa | https://github.com/numba/llvmlite | 01:48:09 |
| hexa | I'm frankly not sure why they lag behind that much | 01:48:24 |
| emily | LLVM does a fair amount of breaking changes I think. | 01:50:39 |
| emily | you're meant to upstream things | 01:51:13 |
| emily | fwiw they apparently have experimental LLVM 16 support | 01:52:53 |
| emily | from the Fedora package:
# For 0.44.0, LLVM 15 is the default. LLVM 16 support is “experimental;” we can
# try it if we need to.
| 01:53:00 |
| hexa | ok, neat | 01:54:33 |
| emily | however a bunch of GHCs keep 15 around anyway so that version is probably relatively safe. | 01:53:43 |
| hexa | the haskell wants to cull some of these | 01:55:23 |
| emily | 12, 13, 14, 16, 18 should be pretty easy drops at this point. maybe 17 too | 01:55:35 |
| emily | plausible that 25.11 will ship with 15, 17, 19, 20 | 01:55:44 |
| hexa | https://discourse.nixos.org/t/nixpkgs-ghc-deprecation-policy-user-feedback-necessary/64153 | 01:56:08 |
| hexa |
Especially, old GHC versions are some of the primary consumers of old LLVM versions in nixpkgs which also puts strain on more than the Haskell maintainers.
| 01:56:28 |
| emily | yes, I know | 01:57:21 |
| emily | I've put maybe 5 hours of my life into arguing about ancient GHCs using ancient LLVMs :P | 01:57:38 |
| emily | glad there is finally some movement on it | 01:57:43 |
| emily | AIUI the next Stackage LTS will be 9.10, so if 25.11 shipped with that then we'd still have 9.6, 9.8, 9.10, 9.12, all but the latter of which use LLVM 15 | 01:59:12 |
| emily | it will only help dropping LLVM 12 | 02:00:15 |
| hexa | automatic bugs against packages would be something | 02:00:47 |
| emily | possible that the version constrains can be loosened or that LLVM support could be dropped from old versions (it is only necessary for weird platforms generally, aarch64-darwin was LLVM only in old releases but I think there is a native backend, I don't know what version added that though) | 02:00:55 |
