LLVM Weekly - #265, January 28th 2019

Welcome to the two hundred and sixty-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 https://llvmweekly.org and pass it on to anyone else you think may be interested. Please send any tips or feedback to [email protected], or @llvmweekly or @asbradbury on Twitter.

News and articles from around the web

Papers have been published for the next C++ standards meeting.

On the mailing lists

• Vedant Kumar shares an update on the status of the hot/cold splitting pass, including some early performance results.

• Johannes Rudolf Doerfert has written an RFC on late OpenMP GPU code 'SPMD-zation'. This involves a new middle-end LLVM IR pass that tries to convert guarded mode code to SPMD mode. After some discussions on-list and IRC, Johannes clarified the idea further.

• Alexey Bader reports that Intel's SYCL compiler and runtime sources are now on GitHub and they intend to start sending patches for review in the coming weeks.

• Craig Topper provides an update on the status of asm-goto support in Clang/LLVM.

• David Chisnall has identified a MemorySSA related regression and is looking for help in tracking it down further.

• Yvan Roux has posted an RFC on adding Machine Outliner support for ARM.

LLVM commits

• The agreed toolchain update policy is now documented in the LLVM Developer Policy. r351765.

• atomicrmw now supports fadd and fsub. r351850, r351851.

• ADT headers no longer use __attribute__((always_inline). Instead, people are recommended to compile a Release build in order to get better performance executing the unit tests. r351891.

• MS SEH extensions __try/__except/__finally are now supported for AArch64 COFF. r351370.

• ScalarEvolution now computes expression sizes estimates. r351725.

• The LangRef description of volatiles has been clarified. r351772.

• HWASan memory access checks are now emitted using small outlined functions on AArch64, meaning there is much less code size bloat but slightly higher performance overhead. r351920.

• HotColdSplit is now run earlier in the pipeline, resulting in substantial code size benefits. See the commit message for full details. r352080.

• i32 variable shifts and sdiv/udiv/urem are now custom-legalised to target-specific nodes. This results in a more reliable selection of the relevant *W instruction (e.g. SLLW, DIVUW etc). r352169, r352171.

• A new TableGen language feature !cond was added, providing an alternative to nested !f. r351285.

• HotColdSplit gained a new cost model to control splitting behaviour. r352228.

Clang commits

• clangd gained an IncludeFixer which can suggest include insertion fixes when e.g. there is a member access into an incomplete type. r352361.

• _Float16 is now disabled by default for targets other than ARM and SPIR. r352221.

• clang-analyzer checkers now declare their dependencies. r352287.

Other project commits

• LLD performance for Identical Code Folding was improved. r351899.

• The LLDB build system was updated to use the host llvm-tblgen binary when cross-compiling. r352235.