Radeon GCN Compiler Back-End Might Still Be Merged For GCC 9

The AMD Radeon "GCN" compiler back-end to the GCC open-source code compiler might still be merged ahead of the GCC 9 stable release due out in April.

Over the past year CodeSourcery / Mentor Graphics has been working on the AMD GCN back-end to GCC so that OpenMP/OpenACC code-bases can run on the latest Radeon/FirePro GPUs using this common C/C++ compiler. While right now there is just the "AMDGPU" LLVM compiler back-end, this work would finally open up OpenACC/OpenMP GPU offloading for the GCC compiler with recent Radeon graphics cards.

In recent months they have revised their patches in hopes of getting the new GPU target approved for adding to GCC 9, the next annual release for the GNU Compiler Collection.

While GCC 9 is now under its final stage of development focused on just regression fixes and documentation updates, this Radeon GCN back-end will still be permitted to land.

Longtime GCC developer Richard Biener confirmed that it's still permitted for landing during this final stage of development once receiving approval on the new code. This is allowed since it's just new code being introduced and doesn't risk regressing existing targets. Though at this stage it just looks to be the AMD GCN code itself and not the needed changes to OpenMP/OpenACC offloading, which would then need to wait for next year's GCC 10.

For those wanting GCC-based Radeon compiler support right now, Mentor's CodeBench Lite supports GPU offloading and is based on the same code as what's being worked on for upstreaming.