LLVM Weekly - #619, November 10th 2025
Welcome to the six hundred and nineteenth 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
Arm engineers blogged about their contributions to LLVM 21.
The next Cambridge (UK) LLVM social will take place on November 27th.
The next LLVM Meetup Darmstadt will take place on Nov 26th.
According to the LLVM Calendar in the coming week there will be the following.
- Office hours with the following hosts: Aaron Ballman, Johannes Doerfert
- Online sync-ups on the following topics: Flang, MLIR C/C++ frontend, alias analysis, modules, GlobalISel, lifetime safety, LLVM/Offload, SPIR-V, HLSL.
- For more details see the LLVM calendar, getting involved documentation on online sync ups and office hours.
On the forums
Peter Rong posted an RFC on providing open evaluation and better onboarding for MLGO (Machine Learning-Guided Optimization), reflecting on the various challenges people face when trying to use or evaluate things like MLGO for register allocation.
Slides are now available from the GPU/Offloading LLVM workshop. Additionally, notes have been posted from the Embedded Toolchain Workshop.
Krzysztof Parzyszek posted a Flang RFC that proposes breaking up compound OpenMP directives in the AST intead of doing it during lowering to MLIR.
Matthias Springer suggests adding to MLIR the ability to execute operations with unsupported floatoing point types by calling into APFloat. The primary motivation is for testing, and the proposal has had a lot of support so far.
Marc Auberer started a discussion about adding new end to end LLVM IR test cases.
Farzon Lotfi made a Clang RFC proposal related to supporting the swizzle-style suffix access syntax for HLSL matrix types.
Naveen Seth Hanig proposed breaking clangFrontend’s dependency on clangDriver by adding a clangOptions library.
Fangrui Song shared thoughts on a compact section header table for ELF.
Jan Svoboda started an RFC discussion on having Clang use a file system sandbox by default for developers builds, which would enforce the use of
vfs::FileSystemand disallow directllvm::sys:fscalls.Luke Lau notes that nightly RISC-V performance numbers are now being reported via LNT for the SiFive P550.
LLVM 21.1.5 was released.
Tarun Prabhu suggests that Flang should accept and ignore gfortran specific options, printing a warning.
LLVM commits
The DFAJumpThread pass was enabled by default. This is reported to give a 13% performance improvement on CoreMark on X86. 0ba7bfc.
The AArch64 backend got better at coalescing stack adjustments in the prologue or epilogue in the presence of SVE. 33609bd.
Work on LLVM’s content addressable storage (CAS) utilities continued with the addition of
UnifiedOnDiskCacheandOnDiskCASas well as anllvm-cascommand-line tool for debugging. 6747ea0, ebb61a5.llvm-dwarfdump learned the
--child-tagsoption to filter by the given DWARF tags. f8656ed.The
nocreateundeforpoisonIR attribute was introduced. f037f41.The
llvm.reloc.noneintrinsic was added, and will emit a no-op relocation against a given operand symbol. 5f08fb4.llc now has a
--save-statsoption which works similarly to Clang’s option of the same name. 96a5289.
Clang commits
C intrinsics were added for the SiFive RISC-V XSfvfexp and XSfvfexpa extensions. 4cb8f97.
A CMake example for linking against libclang was added to the documentation. 71022d1.
counted_byis now allowed onvoid*as a GNU extension. f29955a.
Other project commits
compiler-rt now has a build option for execute-only code generation on AArch64. 9d18e92.
printf error handling was added to LLVM’s libc. 9e2f73f.
LLDB gained
SBFrameListfor iterating over stack frames lazily. d584d00.The XeGPUOptimizeBlockLoads pass was added to MLIR. 9703bda.
The new AccImplicitData MLIR pass will automatically generate data clause operations (e.g. copyin, copyout) for variables used within OpenACC compute constructs that do not already have explicit data clauses. 28c6ed5.
The MLIR pygments lexer was improved. 7ac6a95.