LLVM Weekly - #666, October 5th 2026
Welcome to the six hundred and sixty-sixth 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
Mohamed Emad wrote on the LLVM blog about their GSoC project to share LLVM-libc’s floating-point routines with compiler-rt.
The EDG C++ frontend has now been open sourced.
The deadline for registering for the 2026 LLVM Developers Meeting is on October 16th.
The next Portland LLVM social is taking place on October 22nd.
The next Bay Area LLVM meetup will take place on October 26th.
According to the LLVM Calendar in the coming week there will be the following:
- Office hours with the following hosts: Quentin Colombet.
- Online sync-ups on the following topics: formal specification, MLIR C/C++ frontend, pointer authentication, Clang static analysis, Clang C and C++ language working group, Flang, RISC-V, embedded toolchains, HLSL, qualification, LLVM libc math library.
- For more details see the LLVM calendar, getting involved documentation on online sync ups and office hours.
On the forums
Bruno Cardoso Lopes shared the September ClangIR progress report. 40 out of 47 SPEC CPU 2026 benchmarks compile and pass, 2 fail to link, and 5 have runtime failures (miscompiles).
Nikita Popov proposed a change to LLVM’s AI tool policy, to forbid AI-generated communication, aiming to prevent people submitting PR descriptions and RFCs written by AI tools.
Tom Stellard shared a PSA on project-wide issues we have been facing with merging failing to work on GitHub. It’s a very frustrating situation. Hopefully GitHub will fix it soon.
Mitch Briles started an RFC discussion on respecting freeze semantics in backends. “The SelectionDAG, FastISel, and GlobalISel selectors on AArch64, AMDGPU, RISC-V, and X86 all unconditionally lower freeze/ISD::FREEZE/G_FREEZE to a COPY. If the input to the freeze is undef or poison, what’s left is a COPY of IMPLICIT_DEF, which represents an undef value. ProcessImplicitDefs and the register coalescer then treat the COPY as undefined, and different uses can read different registers. The consequence is a long history of miscompiles”.
Yaxun (Sam) Liu posted an RFC on adding an opt-in option to llvm-cov to exclude annotated source lines and ranges from coverage reports.
Kunal Chadha would like to add a new attribute that indicates the byte range written to by a function for a pointer argument.
LLVM commits
Mips support was added to JITLink. 2820795.
Python shebangs were all moved to
#!/usr/bin/env python3. 82f8753.IRBuilder was updated to always produce canonical constexpr GEPs in ptradd form. e65dba4.
CompressInstEmitternow supportsHwModeresgisters. This allows using a single instruction definition for compressed instructions like RISC-V C_ADDI4SPN/C_ADDI16SP across HwModes instead of duplicating the instruction definitions for each mode. 779a0f493.AMDGPU support was added to LLVM’s ABI library. cd94d18.
The first part of the RFC to declare library command line options in tablegen was committed. 17fdd96.
LoopAccessAnalysis gained stencil group merging to reduce runtime pointer checks. 70a0078.
LLVM IR now supports alias scope domains with disjoint scopes. c851fb6.
Clang commits
As part of reflection support,
std::meta::infois supported for primitive types. 101f575.-fcopyprofwas added to Clang to use the copy profiler. 08d36e8.A one-entry cache was added to
ASTConext::getDeclAttrs()which provides a small but measurable speedup when building large projects like Clang. 4fce7d6.__builtin_sort_packwas implemented. a822257.