LLVM Weekly - #416, December 20th 2021

Welcome to the four hundred and sixteenth 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

Rui Ueyama has released version 1.0 of the mold linker. It aims to be a much faster drop-in replacement for linkers like LLD and GNU gold. On a related note, MaskRay blogged about why ld.lld isn't faster than it is.

On the mailing lists

• Tom Stellard shared an update on the 13.0.1 release. Backport requests will be accepted until Monday December 20th, and rc2 will be tagged shortly after.

• Nikita Popov raises the question of further specifying the interaction of sret parameters and unwinding, with the goal of enabling more optimisations.

• Tom Stellard kicked off an RFC discussion on introducing a new automated release workflow using GitHub Issues and Pull Requests. Respondents asked about the impact in terms of email notifications.

• Weining Lu shared an RFC on adding a new backend for the LoongArch ISA. The work done so far is public, and is being re-factored into small incremental patches as was done for the RISC-V, CSKY, and VE backends.

• Chris Sears started a discussion on tidy-llvm.py, a tool to help run clang-tidy on LLVM itself. Some respondents suggested using a compilation database as a workaround.

• James Nagurne is seeking guidance, examples, or docs related to “super passes”.

• LLVM GPU News #25 is out.

LLVM commits

• Prologue and epilogue generation was implemented for z/OS target. 3f1c403.

• Details regarding MemorySSA’s precision were documented. ac994f8.

• The location description DWARF extension, one of the DWARF Extensions for Heterogenous Debugging used by the AMDGPU target, was documented. c6be2ad.

• Named opcodes are now supported in the .insn directive for RISC-V. 10a7198.

• The llvm-debuginfod-find tool was added. It wraps the debuginfod library to query debuginfod servers for artefacts. f0ca8d2.

• Instruction selection support was added for STRICT_{FADD,FSUB,FMUL,FDIV} on RISC-V. 3926893.

• Deferred inlining was disabled, as with the new pass manager it has been observed to have instances of ‘catastrophic inlining’ (see the detailed commit message for more details). a8c2ba1.

• The process for working with patch series on Phabricator was documented. 2dbe589.

• The RegAllocEvicationAdvisorAnalysis pass was introduced. 0910380.

• The constant index type for extractlement/insertelement is now canonicalised to i64. e6ad9ef.

Clang commits

• The CLANG_DEFAULT_PIE_ON_LINUX CMake option was added to emulate GCC’s --enable-default-pie. 1042de9.

• Support was implemented for P2085, out-of-class defaulting of comparison operators. 5fbe21a.

• Warnings can now be disabled for integer to floating-point narrowing conversions for cppcoreguidelines-narrowing-conversions. 9198d04.

Other project commits

• The v3 ThreadSanitizer runtime was re-committed and enabled by default. b332134, 9fb8058.

• Microsoft’s floating-point to_chars implementation was contributed to libcxx. It uses the Ryu and Ryu Printf algorithms. abb5dd6e.

• A FreeBSDKernel plugin was added to LLDB that supports reading FreeBSD kernel core dumps. 9c7fbc3.

• Patches implementing a new frontend pattern language for MLIR’s PDL have started to land. See River Riddle’s presentation at a recent MLIR open meeting for more information. 11d26bd, 322691a, d7e7fdf, and more.

• A memory allocation rewrite pass was added to Flang, that can reclassify memory allocations from stack to heap when appropriate. 3d092e3.

• LLVM’s libc gained a strtold implementation. aa1902f.

• Documentation was added on how to add a math function to LLVM libc. a2b3e6b.

• P0898R8 (monadic operations for std::optional) was implemented for libcxx. 17cfc57.

• In the ELF LLD linker, scanRelocations was split into scanRelocations and postScanRelocations. Postponing some work to postScanRelocations gives more flexibility, as described in the commit message. cf783be.

• LLDB gained the ability to cache symbol tables for subsequent debug sessions can start faster. da816ca.

• Reproducer replay functionality was removed from LLDB. fa12606.

• Scalable vector support was added to MLIR’s VectorType (meaning a number of operations can be removed from the ArmSVE MLIR dialect). a4830d1.