LLVM Weekly - #196, Oct 2nd 2017

Welcome to the one hundred and ninety-sixth 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

Video and slides are now available from the talk "Getting Started with LLVM. The TL;DR Version", by Diana Picus.

If you're interested RISC-V support for LLVM, you might appreciate my recent RISC-V LLVM status update on the lowRISC blog. See below for details about the upcoming birds of a feather session and a working session after the LLVM dev meeting.

Reminder: the LLVM Dev Meeting is now just a couple of weeks away (Oct 18-19th, San Jose). You can see the whole program on sched.com. Registration will be closing soon. Additionally, Tanya is looking for volunteer session moderators.

Videos from CppCon 2017 are now appearing on YouTube.

On the mailing lists

• I (Alex Bradbury) am collecting topics for the RISC-V birds of a feather session at the LLVM Dev Meeting. This will be held Wed 18th Oct, 4.20-5.05pm. I'm also organising a longer working session, for Friday 20th October, which will also be held in San Jose. See the linked thread for more details and please contact me at [email protected] if you would like to attend.

• Tom Stellard has shared the proposed release schedule for LLVM 5.0.1.

• Discussion on the RFC for an IR-level interprocedural code outliner has continued. Jessica Paquette has proposed how code could be shared between the MachineOutliner and this IR-level outliner.

• Brian M. Rzycki has proposed adding new APIs for "CFG-level" IR manipulation. The hope is this will reduce code complexity in the common case. Daniel Berlin responded with some feedback on this direction.

• Joerg Sonnenberger has written on RFC on enabling reproducible builds using Clang. New command-line options are needed prevent absolute paths leaking in to builds.

• Davide Italiano has proposed adding an intrinsic for 'cls' (count leading sign bits). This is motivated by making it easier to generate the AArch64 cls instruction.

Review corner

The LLVM Weekly review corner serves to highlight patches that are stuck waiting awaiting review, or work from first-time contributors. See here for more information and how to submit you work for inclusion. Of course the hope is that highlighting these patches will enable LLVM Weekly readers will step up and help to get them merged. I'll be reporting back each week on any activity generated on these patches, as well as sharing a new batch. If you want your patch included you must submit it via the linked form.

Last week featured four patches, with three out of four seeing fresh review activity after inclusion. D36357 sadly still has no feedback. Thanks to Dimitry Andric, Matthias Braun, and Anna Zaks for contributing review comments to last week's patches.

Unfortunately I've received no valid submissions for inclusion in this week's issue. I'm going to keep the review corner trial going for a little longer, but obviously without patch submissions we might conclude that the review corner experiment should go to an end.

In the absence of external submissions, let me advertise my own stuck review:

• "Adds initial code generation support for RV32 (32-bit RISC-V). Along with D29934, this is blocking three successfully reviewed patches." D29933, patch by Alex Bradbury.

LLVM commits

• MC-layer support for AArch32 v8.3-a complex numbers has been added. r314511.

• The X86 backend now uses INSERT_SUBREG rather than SUBREG_TO_REG when matching an anyext GR32 pattern. r314152.

• The MachineOutliner has been updated to avoid saving restoring the link register if possible when used with the AArch64 target. r314341.

• Support has been commited for common fixups and relocations in the RISC-V MC layer. r314389.

• TargetInstrInfo::getInlineAsmLength has learned to parse certain forms of the .space directive. This is primarily useful for crafting testing inputs that exercise the branch relaxation pass. r314393.

Clang commits

• Design documentation has been committed for Clang's refactoring engine. r314509.

• A new "vexing parse" warning has been added for when there is ambiguity between a variable declaration and a function-style cast. r314570.

Other project commits

• LLD gained a huge speedup for -r and --emit-relocs, which reduces link time of the Linux kernel from 12.95s to 0.55s. r314282.

• LLD's string merging has been parallelized. This reduces the link time of a Clang debug build from 7.11s to 5.16s (4x faster than gold, as tested by the committer). r314588.

• The benchmark numbers in the LLD documentation have been updated. r314617.

• The sanitizer allocator's "release memory to OS" implementation has been rewritten. See the commit message for full design details. r314311.

• Polly gained a short document describing Polly's performance for GEMM (generalized matrix multplication), and the impact on compile time. r314419.