At the recent European LLVM meeting in Paris, Andrey Bokhanko and Alexey Bataev of Intel covered their work on supporting OpenMP within LLVM.
LLVM's lack of OpenMP support for multi-processing is what's made the Clang compiler not too suitable at the moment for some multi-threaded benchmarks, as Phoronix tests so frequently show. For years now there's been efforts to support OpenMP in LLVM/Clang, but nothing that's yet been merged to mainline.
In 2012, work on OpenMP in LLVM heated up
but still with the forthcoming LLVM 3.3 release there is not full support for Open Multi-Processing.
In the Intel OpenMP LLVM presentation from France, the state and current work was covered. The slides can be found here
Intel's efforts build upon earlier work done by AMD in supporting OpenMP and the Clang compiler can generate OpenMP RTL calls without changes to the LLVM IR. The GCC compiler uses libgomp as its OpenMP library but with that being GPLv3 and is incompatible with LLVM's more permissive BSD-style license, Intel stepped up to the place. Released in April was the Intel OpenMP runtime that is under a three-clause BSD license. Intel's OpenMP runtime that they put out under the BSD license is their production runtime used by the ICC compiler and Ifort. This Intel OpenMP runtime has been in development since before OpenMP was officially conceived and has proven to be highly-scalable.
What's been implemented and committed so far when it comes to the LLVM/Clang OpenMP support is handling the -fopenmp
compiler switch, supporting the #pragma omp threadprimitive
, and parsing and semantic analysis with AST representation. There's code currently under review for supporting all OpenMP pragmas while still under development is code generation support for OpenMP constructs.
While no time-line was shared in regards to mainline OpenMP LLVM/Clang support, hopefully with the current pace of activity, that there will be some functioning support prior to year's end.