[email protected]

Maintainer of LLVM MC, Clang driver, lld/ELF, and libunwind

[[email protected] @HaskRay]

@jian_w3ng 好在我们那年多了B类(还有C类等)的说法，否则4个男选手名额我就挂了T_T

[[email protected] @HaskRay]

Since June, lld has been getting AI-drafted bug reports in batches. Since late October they end with "a possible fix", and some rando turns each into an AI-assisted PR: the suggested fix verbatim, the reproducer pasted in as the test. Review lands on the maintainer twice.

[[email protected] @HaskRay]

@Sergey_Rura I exported comments as an XML file back in Feb 2025 but never got around converting them to Giscus. LLM tools made it happen.

[[email protected] @HaskRay]

https://t.co/BfqvnqOd1q "Unfortunately comments from Disqus have not been migrated yet" => I imported the comments from the Disqus export (`comments.xml`) with an Opus 5.5-written Python script that calls the GitHub GraphQL API via `gh api graphql`.

[[email protected] @HaskRay]

lld/ELF with my parallel symbol interning patches, vs mold (2026-10-21, Rust), vs wild (2026-11-04), linking ELF clang-release, on Apple M4 Max.

[[email protected] @HaskRay]

lld 24.x optimization preview: parallel symbol interning. Linking clang-release at --threads=8: 181 -> 151 ms

[[email protected] @HaskRay]

lld linking Chromium (CREL, --icf=all): 3.46 -> 2.08 s. In chromium-debug, Moore's partition refinement algorithm performs poorly.

[[email protected] @HaskRay]

Google's requirements have drifted away from the already very broad range that lld, and llvm-project at large, serves. A feature only one company's internal pipeline might need should prove itself there before it lands upstream, where everyone maintains it.

[[email protected] @HaskRay]

In 2020, reviewing lld's --optimize-bb-jumps (https://t.co/sD35KLTMJu), I asked that the pass run inside finalizeAddressDependentContent "to accommodate thunks and linker script symbol changes." The author replied: "Could we do this in a later patch?" There was no later patch.

[[email protected] @HaskRay]

and another lld maintainer independently flagged the misplacement and found my 2020 comment.

When I left Google in 2024, --optimize-bb-jumps still had no users. The answer was always "we will use it." Today nothing in llvm-project passes the flag; only three lld tests do.

[[email protected] @HaskRay]

I was junior; he was far more senior. Landing a feature on a promise of future use, and deferring the reviewer's design concern to "a later patch", is not how open source should work.

[[email protected] @HaskRay]

That placement was a wrong-code bug from day one: the pass shrinks .text after .relr.dyn is computed, so RELR entries point at the wrong words.
Since a change this year, .eh_frame_hdr search keys go stale too.

An AI-assisted patch surfaced it recently,

[[email protected] @HaskRay]

@Temporary111x The Zola+tabi migration is vibecoded:) The website was stuck using
https://t.co/tDfcgZUYaV for 10+ years...

[[email protected] @HaskRay]

https://t.co/gOAU1Mh1j6 Who pays for Clang
Another AI slop post. Commits by employer, What each funder works on, Review, Shared and vendor-specific work

[[email protected] @HaskRay]

Comment on "RFC: Enable ClangIR Build By Default-" https://t.co/BghNIuldyI Current feeling: it might be a huge opportunity cost

[[email protected] @HaskRay]

I spent many hours driving the agent through Discourse threads, slides, source trees, and LinkedIn profiles to generate this AI slop: https://t.co/fwtUrRQD1O ClangIR and the cost of MLIR

[[email protected] @HaskRay]

For C++ programmers who were Clang's main constituency, CIR has to earn its place by being better at what the AST, the CFG, and classic CodeGen do today: diagnostics, safety analyses, and optimizations that need C++ semantics, without making the compiler slower to build or run

[[email protected] @HaskRay]

The new funders build GPU and accelerator compilers, and their other compilers (Flang, Triton, IREE, their own stacks) are built on MLIR. For them, a Clang that emits an MLIR dialect is the point, not an implementation detail.

[[email protected] @HaskRay]

https://t.co/WnMmvk6wup
Indexed forwarding functions such as std::make_unique<A>(...) and std::vector<A>::emplace_back(...).
Hover on an auto variable in a function template shows the deduced type when the template has exactly one instantiation.
Adapt to LLVM 22, 23 and 24.

[[email protected] @HaskRay]

edgls is a C/C++ language server built on the EDG C/C++ front end. It is a fork of ccls with the clang-based parts replaced: the indexer, diagnostics, compile command handling, code completion, signature help, and formatting. https://t.co/cTkRjMA5rq

Verdict: Slower than clang