LLVM 12 Ends Feature Work With Better C++20 Support To Intel Sapphire Rapids + AMD Zen 3

Feature development on LLVM 12.0 has ended along with associated sub-projects like Clang and libc++. Feature work now shifts to LLVM 13.0 while the LLVM 12 stable release should be out in just over one month's time.

LLVM release manager Tom Stellard announced on Tuesday night that the 12.x release branch has been created and the main development trunk open to LLVM 13.0 feature work.

LLVM 12.0-RC1 is expected to be tagged later today while a second release candidate is expected before the end of February. If all goes well, LLVM 12.0.0 will be out around the start of March.

Among the changes that have accumulated for LLVM/Clang 12.0 over the past half-year include:

- Support for the x86-64-v[234] micro-architecture feature levels in matching the behavior of GCC.

- Support for Intel Alder Lake and Sapphire Rapids with -march=alderlake and -march=sapphirerapids, respectively. There is also support for the HRESET, UINTR, and AVXVNNI instructions.

- Support for AMD Zen 3 CPUs with -march=znver3 though this is only the initial support with the refined implementation still being worked on.

- Support for using the "-mtune=" option on x86/x86_64 targets now to request microarchitectural optimizations independent of "-march=". If mtune is not specified, it's still the same behavior as previously with -march= implying the -mtune= value.

- Wrapping up various lingering bits of C++20 language support, like the likelihood attributes. The libc++ library has also worked more on its C++20 support and even starting on C++2b work.

- Much lower clangd memory usage.

- Various AMDGPU back-end improvements.

See more within our past LLVM 12 and Clang 12 articles with a more extensive overview coming in due course. LLVM/Clang 12.0 benchmarks will also be coming soon on Phoronix.