LLVM Weekly - #411, November 15th 2021

Welcome to the four hundred and eleventh 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 http://llvmweekly.org and pass it on to anyone else you think may be interested. Please send any tips or feedback to asb@asbradbury.org, or @llvmweekly or @asbradbury on Twitter.

News and articles from around the web

The 2021 LLVM Developers' Meeting is taking place virtually this week, Nov 16th-19th.

Simon Byrne blogged about dangers posed by ‘fast-math’ compiler flags.

On the mailing lists

• Akash Kothari (on behalf of a consortium of authors) shared a proposal for TLX: Tensor LLVM eXtensions.

• Jeremy Morse posted an RFC on enabling ‘instruction referencing’ variable locations for x86_64. Respondents so far are in favour.

• Johannes Doerfert asked if it would make sense to have a ‘GPU working group’ with regular meetings.

• Mircea Trofin shared an RFC on applying machine learning to live range eviction for the ‘greedy’ register allocator, noting a 0.3%-1.5% improvement in queries per second on selected datacenter-type applications.

• Zarko Todorovski is proposing to rename llvm-cov option -name-whitelist to -name-allow-list.

• Florian Hahn provided an and updated on adding new vector math builtins to Clang.

• Itay Bookstein is seeking feedback on the IR-level semantics of a GlobalIFunc whose resolver is a declaration.

• After some systematic testing with a simple Python script, Jay Foad highlighted some missed optimisations of bitwise expressions. ‘Beginner’ tagged bugs have now been filed.

• Johannes Doerfert posted an updated RFC on introducing non-capturing stores.

• Alan Phipps kicked off an RFC thread on source-based modified condition/decision coverage.

• ORC JIT Weekly #44 is out, providing an update on support for MachO debugging.

• LLVM GPU News #23 is out.

• Arthur Eubanks initiated a thread on removing optimisation size level from the LLVM IR pipeline and relying on minsize/optsize attributes. This generated a fair amount of discussion, both about this specific proposal and about when function attributes should be used to represent optimisation information.

LLVM commits

• The llvm-tli-checker tool was committed. This tool compares TargetLibraryInfo’s opinion of the availability of library function calls against the functions actually exported by a specified set of libraries. 38be8f4.

• MC layer support for CSKY’s 16-bit instructions was added. 4fb282f.

• The initial implementation of a module level inliner was committed. 6cad45d.

• Minimal support for the D programming language was added to LLVM’s demangler. A fuzzer for it was also contributed. 22a1aa5, 9af467e.

• The operands-skip pass was added to llvm-reduce. This will try to remove instructions in the middle of dependency chains. c15f930.

Clang commits

• LambdaCapture-related matchers were added. c0b298f.

• A new clang-tidy check was added to detect misleading unicode identifiers. a35efc4.

• Clang support was added for the SPIR-V target. a10a69f.

• The # NNN-based line control extension was documented. bf6986d.

• Some of the generated resource headers were conditionalised upon the relevant target being registered. This was motivated by the size the RISC-V headers add to the resource directory. c17d9b4.

• A new clang-tidy check was added for the initialisation of absl::Cleanup. 16b07c8.

Other project commits

• A libMLIR-C.so shared library can now be built. c265170.

• compiler-rt’s memprof gained support for a raw binary profile serialisation format. 545866c.

• The performance of string to float conversion in LLVM’s libc was improved. 8298424, 499ca80.

• Outdated LLD design documentation was removed. 43bb5f0.

• Identical code folding (ICF) for the LLD MachO linker can now deduplicate functions with identical unwind info. d9b6f7e.