What's Exciting About LLVM 3.0 & The New Clang

Version 3.0 of LLVM is expected to be released tomorrow, along with a major update to Clang. Here's some of the noteworthy enhancements.

LLVM 3.0 with the adjoining Clang update is the first major update to the Low-Level Virtual Machine since the LLVM 2.9 release last April. LLVM 3.0 was scheduled for a November release (but it was delayed slightly) and marks the point of deprecating LLVM-GCC in favor of DragonEgg, which allows for LLVM optimizers to be used with the mainline GCC compiler front-end via a unique plug-in. Other interesting changes for LLVM 3.0 are listed below.

Clang C/C++ Compiler In LLVM 3.0:

- Much-improved C++ support in building software, the stability of the built software, and the diagnostics of any C++ issues.

- Improved C++0x (C++ 2011) support, but it's still not feature complete.

- Support for some features in the upcoming C1x standard.

- Improved Objective-C support.

DragonEgg For LLVM 3.0:

- GCC 4.6 is now fully supported by this plug-in to replace the GCC optimizers with those from LLVM.

- DragonEgg no longer requires a patched GCC but the plug-in can be cleanly loaded from GCC mainline versions 4.5/4.6.

- The option that allows both the GCC and LLVM optimizers to be used (rather than outright replacing the GCC optimizers with those of LLVM). This option is designed to deliver the "ultimate performance" but is still deemed experimental. (I'll be benchmarking out the new DragonEgg soon.)

- C, C++, Fortran, and ADA remain the officially supported languages for this GCC-front-end with LLVM-back-end plug-in. There's partial support for Java, Objective-C, Objective-C++, and Google Go.

Other LLVM Changes:

- libc++ (the new C++ standard library) and compiler_rt are now dual-licensed under the MIT and UIUC licenses. The LLVM C++ standard library is expected to be the default STL implementation in FreeBSD 10.0.

- VMKit (A Java Virtual Machine for LLVM) now supports pre-compilation, customized virtual methods, better inlining support, and a new exception model.

- LLVM IR now has full support for atomic memory operations as needed to support C++11 (C++0x) and C1x.

- A new "greedy" register allocator replaces the linear-scan register allocator.

- There's significant improvements to the MIPS back-end in LLVM 3.0.

- The MIPS architecture back-end to LLVM has been greatly improved.

- AVX (version 1) is now supported by LLVM, this is the Advanced Vector Extensions as introduced in the AMD Bulldozer and Intel Sandy Bridge CPUs.

- The Assembler/Disassembler has a range of new instructions for Intel Atom, Intel Ivy Bridge, SSE4a/BMI, rdrand, and others.

- Significant performance improvements for ARM Cortex-A9 SOCs. There's also improved Cortex-M series support and performance improvements for ARM NEON code.

More details about the imminent release of LLVM 3.0 can be found on the release notes page. I have already published some LLVM/Clang 3.0 benchmarks but more will be on the way from a greater range of systems and also in benchmarking the DragonEgg plug-in.