LLVM Weekly - #672, November 16th 2026
Welcome to the six hundred and seventy-second issue of LLVM Weekly, a weekly newsletter (published every Monday) covering developments in LLVM, Clang, and related projects. LLVM Weekly is brought to you by Alex Bradbury. Subscribe to future issues at https://llvmweekly.org and pass it on to anyone else you think may be interested. Please send any tips or feedback via email: asb@asbradbury.org, or Mastodon: @llvmweekly@fosstodon.org / @asb@fosstodon.org, or Bluesky: @llvmweekly.org / @asbradbury.org.
News and articles from around the web and events
Iason Karaprodromidis wrote on the LLVM project blog about their GSoC project to add Fortran support to LLDB.
Save the date for EuroLLVM 2027, taking place April 26-28th in Porto. The call for program committee volunteers is open now.
The Meta engineering blog posted about extending LLDB to debug applications that span CPUs and GPUs.
The website for LLVM libc has had a redesign and now gives completeness metrics for each header.
Recordings are now available from the GNU Tools Cauldron 2026.
The next Israel LLVM Meetup will take place on November 24th.
According to the LLVM Calendar in the coming week there will be the following:
- Office hours with the following hosts: Phoebe Wang, Clang maintainers.
- Online sync-ups on the following topics: formal specification, pointer authentication, lifetime safety, vectoriser improvements, Flang, RISC-V, LLVM libc, HLSL, LLVM libc math.
- For more details see the LLVM calendar, getting involved documentation on online sync ups and office hours.
On the forums
Eugen Zalenko proposes to limit the rate of AI/bot-assisted issue creation.
Aaron Ballman is interested in starting a community floating-point working group.
Round table notes are now up for compact unwind on ELF.
Louis Dionne provided an update on work on LNT v5.
Harrison Hao posted an RFC on deprecating and removing support for the R600/TeraScale, having confirmed that Mesa no longer depends on it.
Matthias Springer proposed a new ub.assume operation in MLIR.
Slava Zakharin posted an RFC (and linked a design document) on representing Fortran aliasing rules in LLVM IR.
Andrzej Warzyński and Vicky Nguyen proposed extending update_cc_test_checks.py to support tests that verify multiple IRs within a single source file.
Madhur Amilkanthwar, Sjoerd Meijer, Sebastian Pop, Elvina Yakubova, and Yafet Beyene shared an RFC on communicating array extents to LLVM using llvm.assume.chained. “Fortran guarantees that every array subscript stays within its dimension, but the guarantee is lost when Flang lowers an access to a linear address. We propose llvm.assume.chained, a pointer-forwarding intrinsic whose operand bundles carry an array’s extents and strides from Flang to delinearization and dependence analysis. Accesses use the returned pointer as their base, so they retain the array information when passes rebuild loads and stores. We also implemented the same facts as bundles on the existing llvm.assume, one per access: both reach the same loop interchanges and runtime gains, but the chained intrinsic does so at a lower compile-time cost. We compare the two, along with instruction metadata.”
Alina Sbirlea would like to host LLVM formal verification tools as part of the LLVM project. i.e. a central repository of shared input test cases paired with expected tool outputs. There was some discussion about how useful this collection of test cases would be, vs bringing in more tools to the monorepo, vs providing a compiler explorer instance.
Andrzej Warzyński and Federico Bruzzone are seeking to align MLIR CODEOWNERS with those who are actively reviewing.
Kay Hicketts and Valentyn Yukhymenko shared an RFC on user-defined annotations for static analysis. This would provide an annotation mechanism that lets library owners describe how their APIs should be interpreted by tools such as
clang-tidy.Md Abdullah Shahneous Bari started a discussion on changing the handling of fast-math flags on MLIR linalg and vector ops.
Caleb Zulawski kicked off an RFC thread on adding an intrinsic to query enabled target features after inlining. Nikita Popov clarified this would be used for the compile-time detection case, for situations where runtime detection is too expensive.
LLVM commits
An
llvm.vector.shuffleintrinsic was added. 47e4e82.The IRBuilder now supports GEP canonicalisation, where arbitrary GEPs are converted into a sequence of ptradd or scaled ptradd. It was additionally enabled by default. ede30f6, 847ab25.
The LangRef data race definition was relaxed. 4687e1e.
The way backends handle undefined values was documented. df0a91f.
MC layer support, intrinsics, builtins, and instruction selection support was added for the Hexagon HLX (eLementwise eXtension) instructions. 23df5cc, 96e9c10.
MC layer support was added for the Hexagon Matrix eXtensions (HMX). 34f2fb6.
optnonenow impliesnoinline. 473f6ad.
Clang commits
Clang now supports multiple in-process cc1 jobs in a
ToolSession. This can be used to support multi-file compilation in the browser in Wasm. cbf54d6.RISC-V XSfmm ABI attributes are supported, which describe whether a function reads/writes matrix state. 09dcbfb.