LLVM Weekly - #333, May 18th 2020

Welcome to the three hundred and thirty-third 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

It has now been five years since Rust 1.0 was released (which of course, uses LLVM as a backend). Read all about progress since then on the Rust blog.

Egor Bogatov wrote a blog post on implementing a new peephole optimisation in LLVM.

On the mailing lists

• Ties Stuij posted an RFC to collect feedback on next steps for BFloat, linking to patches currently up for review.

• The 7th edition of the MLIR newsletter is now out, higlighting the mlir-npcomp project (prototype for compiling numpy programs) alongside a range of other developments.

• Alexey Lapshin reported on efforts to have LLD remove obsolete debug info. This --gc-debuginfo option has a meaningful impact on size (reducing debug info size by 50%), but increases linking time substantially.

• Richard Barton provided an update on F18/Flang buildbots.

• Eric Christopher is seeking feedback on LLDB pre-merge testing.

• ORC JIT Weekly 16 provides information on JITLink ELF/x86 and updates on removable code.

• Daniel Grumberg posted an RFC on generating CC1 command lines from a CompilerInvocation instances.

• David Spickett proposes changing Clang's tool search path priority so more specific names on the user's PATH are found before less specific names in the clang install dir.

• Simon Moll authored an RFC on extending Clang to allow bool as a valid vector element. Some of the following discussion focused on whether a vector of bool would be represented as a vector of i1 in IR, and whether this may cause problems.

LLVM commits

• The bfloat type was added to LLVM IR, intended to support the BFloat16 type introduced in Armv8.6. 8c24f33.

• The ARM backend gained support for the machine outliner. 0e4827a.

• An x86 Speculative Execution Side Effect Suppression pass was added, to mitigate Load Value Injection. Support for LVI load hardening was also added. bf95cf4, e97a3e5, 8ce078c, 08b8b72.

• The git-svn utilities were removed, as the Git migration makes them redundant. f056dac.

• Constant pools are now handled in the RISC-V load/store peephole optimisation, improving codegen especially for loading FP constants. 969e703.

• A comment directive, COM: was added to FileCheck. a1fd188.

• AMDGPU gained support for spilling the frame pointer to memory. 09253b6.

• Alignment of allocas will no longer be promoted beyond the stack alignment. c9c930a.

• LLVM's DIModule debug metadata was extended to support Fortran modules. e59744f.

• Support was added for the Marvell ThunderX3T110, including an initial definition of the micro-ops for the scheduling model. 382d3a8.

• Callbacks were added to allow overriding the datalayout after parsing it from a file. 4532a50.

Clang commits

• Clang gained support for the matrix type, which can be enabled with -fenable-matrix. 1065869.

• The static analyzer's PlacementNewChecker was extended with a chec for insufficient storage of arrays and bad alignment. 7c37684.

• Use of _ExtInt in inline assembly is now prohibited. 5f1f4a5.

• clang-format now supports formatting of aligned nested conditionals. 4db9409.

Other project commits

• LLVM libc gained implementations of fabs and fabsf. 4a39a33.

• LLDB will no longer disassemble large functions by default. f665e80.