LLVM Weekly - #425, February 21st 2022

Welcome to the four hundred and twenty-fifth 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

A new paper Compiler Support for Sparse Tensor Computations in MLIR discusses adding support to MLIR for generating sparse code automatically from a sparsity-agnostic definition of computation.

Learning Branch Probabilities in Compiler from Datacenter Workloads presents an approach to use machine learning to estimate branch probabilities.

Jonathan Salwan published the VMprotet-devirtualization repo, an approach to devirtualizing pure functions proteted by VMProtect 3.x (lifting the symbolic representations to LLVM IR).

On the forums

• Nikita Popov provided a status update on opaque pointers. It’s now possible to compile most C/C++ code, with or without optimisation, in opaque pointer mode. There’s more work to be done, but the end is in sight. Congratulations and thank you to everyone involved so far.

• Stella Laurenzo proposed a new ‘ml_program’ MLIR dialect, intended to support the common structural components needed to represent compiled programs from popular machine learning frameworks.

• Fangrui Song started a thread to discuss speeding up LLD through parallel input file parsing and the changes that would be needed to support this.

• Tanya Lattner shared notes and next steps from a recent meeting about Outreachy and GSoC internships.

• Matthias Gehre kicked off an RFC discussion on adding support for division of large _BitInt.

• Yuanfang Chen is working on adding an equivalent to MSVC’s “Just My Code” stepping in LLDB and is looking for input from people interested in this feature.

• Anton Korobeynikov created a pinned topic providing guidance on LLVM in GSoC 2022.

• Philip Reames shared an RFC on Adding a SCEVCompareExpre to ScalarEvolution. Unfortunately it’s not had any responses yet, so please do take a look.

• Fangrui Song proposed changing the behaviour of Clang searching for multiarch include and library paths in the presence of --target.

LLVM commits

• The frontend performance tips document was updated to recommend the use of poison values instead of undef and to tag function attributes with noundef where possible. 8608650.

• The ‘uwtable’ attribute was extended with the unwind table kind (sync/async). 6398903.

• A ‘ptrauth’ (pointer authentication) operand bundle was introduced. c703f852c.

• llvm.vp.fptosi.* intrinsics were added. 03e83cc.

• Documentation was updated on opaque pointers and the migration to using them. 4846568.

• MC layer support for the RISC-V ‘Zfinx’ extension (floating point support without a separate FP register file) was added. 7798ecc.

• The AArch64MIPeepholeOpt pass was expanded to cover the SUBS and ADDS instructions. c69af70.

Clang commits

• The memory overhead of #define directives was reduced. For a file with 7.8 million #define directives, the persistent memory allocation footprint was reduced from 5.28GB to 4.07GB. 00cd6c0.

• Support was added for the [[noreturn]] attribute in C2x. 5029dce.

• The disable-noundef-analysis flag was renamed to -[no-]enable-noundef-analysis to better match the naming conventions of other flags. b529744.

Other project commits

• A new llvm-libgcc folder was added, containing files to allow building libunwind as a libgcc “front” that can be used to replace libgcc.a, libgcc_eh.a, and libgcc_s.so (see the commit message for more details). c5a20b5.

• The ‘oilist’ (order-independent list of clauses) primitive was added to MLIR. 5bec1ea.

• The upstreaming of the fir-dev branch to Flang continues, with patches landing to support lowering basic functions with a scalar integer/logical return value, simple scalar assignment, and more. ad40cc1, e641c29.

• A platform-independent buffered file I/O data structure was added to LLVM’s libc. 4ef02da.

• The llvm-libc index page was rewritten. 171cb8f.

• libcxx will now run clang-tidy during CI. f10909a.