Ramkumar Ramachandra

Upstream LLVM contributor. Expect a mix of technical posts interspersed with cute real-world musings.

Public Key

npub1e8evth4kpk2uv9mmf9ywyff9sgg9swgur8fzj5u3t2xvpx3gaggql9s8wf

NIP-05 Address

artagnon@mathstodon-xyz.mostr.pub

Profile Code

nprofile1qqsvnuk9m6mqm9wxzaa5jj8zy5jcyyzc8ywpn53f2wg44rxqng5w5yqpzemhxue69uhhyetvv9ujumn0wd68ytnzv9hxg0wly7u

Publishing to

relay.nostr.band

Author Public Key

npub1e8evth4kpk2uv9mmf9ywyff9sgg9swgur8fzj5u3t2xvpx3gaggql9s8wf

Show more details

Last Notes

2024-12-12 14:25:27

Alive2 now supports X86 vector intrinsics, and this change should be online on Compiler Explorer soon! It's a huge piece of work originally authored by @regehr's PhD students, and we just finished verifying that the #LLVM test suite passes with the change.
https://github.com/AliveToolkit/alive2/pull/1128

2024-11-28 22:56:00

- reply

@nprofile…gyrf Can we organize a chat before the Christmas break (before the 16th of December)? I can think about what we discussed and plan next steps while on vacation.

2024-11-27 17:18:35

- reply

@nprofile…gyrf I’d be happy to collaborate on future work, including restricting the search space, and adding support for RISC-V vectors. However, I’d first like to get familiar with the code, and figure out the maintenance story (should I be sending PRs and commit to upstream maintenance, or maintaining my own fork). For now, there are some outstanding PRs in Minotaur and Alive2, on which the loop needs to be closed.

2024-11-27 16:32:01

- reply

@nprofile…gyrf Good news! My employer is over the moon about this project! I got Minotaur working, with the CI setup pending. Minotaur is a lot slower than I imagined, so I have to think a little more about the setup before getting it to automatically report missed optimizations.

2024-11-22 20:06:10

- reply

@nprofile…gyrf Just finished reading the paper and skimming the code. I think it's small enough, and well within my grasp to maintain and extend Minotaur: the only question is if I will be able to sell this work to my employer, to also get time to do this as part of my day-job. Potentially, that could mean that I'd have time to eventually make my way to contributing to Alive2, and implement much-needed support for those loops we talked about earlier. I'm very optimistic about selling this to my employer though: I think this thing is going to report a huge number of missed optimizations that will look very good. Otherwise, my Sundays are taken up by the simplicial-cubical formalization work we're doing, and I spend the one day, Saturday, not doing work.

2024-11-22 19:27:30

- reply

@nprofile…gyrf I see some open issues: as long as there's some kind of maintainer to accept PRs, I think I'm good to go. If not, I might have to maintain my own fork.

2024-11-22 19:27:30

- reply

@nprofile…gyrf I see some open issues: as long as there's some kind of maintainer to accept PRs, I think I'm good to go. If not, I might have to maintain my own fork.

2024-11-22 19:18:21

- reply

@nprofile…gyrf Haha, it was a bit naive of me to assume that something like this didn't already exist. On second thought, I think I will fork Yingwei's corpus, wire it up to Minotaur, run on CI, and report the missed optimizations. It might just be a weekend project, and I think the effort will be well worth it, especially since Minotaur seems to work with vectors. I will have a never-ending supply of work! Now, if what I plan is already implemented somewhere, do let me know!

2024-11-22 18:56:12

- reply

@nprofile…gyrf Ah, I'm also unaware of Souper and Minotaur. I'll investigate them, and bring them up the next time I have the chance to talk with him. Thanks!

2024-11-22 18:47:46

- reply

@nprofile…gyrf Also very cool: automatically finding missed boolean optimizations with the help of Alive2 (Yingwei's project).
https://github.com/dtcxzyw/llvm-tools/issues.

2024-11-22 18:42:19

- reply

@nprofile…gyrf Oh, do chime in on the review then! I don't know how Florian even conceptualized ConstraintElimination: seems like a very clever idea.

2024-11-22 18:24:56

- reply

@nprofile…gyrf Very nice! Just had the chance to look at this, and Yingwei's patch: it's a significant improvement in real-world code.

2024-11-18 18:52:45

- reply

@nprofile…gyrf 69 function attributes?! On a related note: I recently learnt about flush-denormals-to-zero (DAZ), and incidentally fixed the handling of zeros in GVN, as part of a larger patch. Although it was technically a bug fix, I suppose it’s a bug nobody has hit or cares about. From a correctness perspective, it’s very unsettling to know that that are probably a billion such bugs from the explosion of features. I do sincerely hope that the number of serious real-world bugs is decreasing with time, despite new features being added.

2024-11-18 17:51:25

- reply

@nprofile…gyrf Sometimes, I feel that we need to slow down on the new features (samesign, fmf on more instructions, nusw, errno memory effects, fp literals, ptradd and many more), and focus on fixing the technical debt. Maybe it is mangeable though, as we have a lot of contributors.

2024-11-18 17:19:02

- reply

@nprofile…gyrf Apparently, the long-term plan is to get rid of undef altogether (!).

2024-07-29 12:12:53

- reply

@npub1cm7…32y8 See
https://lore.kernel.org/lkml/20240715-x1e8-zap-name-v2-1-a82488e7f7c1@linaro.org/
https://www.phoronix.com/news/Linux-Disabling-X-Elite-GPU

2024-07-28 16:05:34

So the #Qualcomm #Snapdragon X Elite release still doesn’t mean that we will have usable #Linux laptops anytime soon. Apparently, they’ve cut some shady deals with #Microsoft and other companies to OEM-sign the GPU and provide a non-ACPI boot up process. The #kernel community seems pretty annoyed.

2024-03-29 15:09:03

I recently learnt that #Rust's type system is unsound. There was apparently a huge formal verification effort that failed. Honestly, if you want a sound and insanely advanced type system, #Coq is your best bet.

2024-01-10 12:42:25

In #kernel news, a 2018 thread on #LKML has recently had some activity, notably from H. Peter Anvin. It’s about using #cpp in the kernel. While it might arguably go nowhere, HPA’s lengthy commentary is worth reading.
https://lore.kernel.org/lkml/3465e0c6-f5b2-4c42-95eb-29361481f805@zytor.com/