For a while, SiFive and LLVM were both developing support for RISC-V Vector 1.0. LLVM is now the only one in active development.

Jim Wilson, a GNU developer of decades, works for SiFive and while he wasn’t the one doing the work, it seemed likely he knew who did, so this is authoritative. He recently said:

There is no actively maintained gcc rvv support, and no ongoing gcc rvv development. Current work is all in LLVM, and LLVM is recommended if you want rvv support. … SiFive abandoned the gcc rvv work and is doing only llvm rvv work now. The gcc rvv branch is badly out of date.

Reading deeper into the GCC development list, this is really just a form of tough love as work on GCC’s RISC-V vector (and auto-vectorization in general) has been talked down before in July, 2021: 

 It isn’t up to date with the evolving RVV ISA spec, it isn’t up to date with the evolving RVV intrinsics spec, there are ugly hacks in the vectorizing optimization passes required to make it work, there is no autovectorization support, it is missing basic optimizations like eliminating duplicate vsetvli instructions, etc. The current status is that it is only useful as a toy for demos. SiFive and a few other organizations are contributing to the LLVM vector support, but no one is contributing to the gcc vector support. Alibaba has expressed some interest in contributing recently but it isn’t clear how we will handle their patches yet. The current stuff was mostly done by SiFive, but SiFive is not currently interested in funding this work.

I’m less sure of rjiejie‘s credentials, but they may work for Alibaba/T-Head, the maker of the cores used by Allwinner in D1 and D1s. That may be the “we” in his comment:

We have also supported/maintained the RVV v1.0 feature, you could download prebuilt gcc toolchain from Alibaba website[1].

Registration for the site is required and Google Translate doesn’t handle it well, so I’m not sure, but that may be a path for someone really needing GCC with Vector 1.0. It’s not clear if that work handles 0.7.1 of Vector as was used in D1. The C910-based products also seem to support 0.7.1, so their 1.0 support must be for future chips.

LLVM is one of several projects that has struggled with the issues of handling multiple V versions in the same code, but their resolution wasn’t clear. Simulator QEMU “solved” them problem when adding 1.0 by dropping support for Vector 0.7.1.

SiFive is “only” one of many core vendors and that they can’t be expected to carry the development/maintenance/support for such things by themselves, but it’s surprising (to me) that they’ve halted development.

T-Head has binaries (and maybe source) for GCC that supports V1.0 and probably 0.7.1, though that may be in a branch as it’s pretty clearly a dead end now that V1.0 has been ratified. LLVM and SiFive were, at least at one time, partnering in LLVM development. LLVM seems to have an active plan and are shipping V1.0 support now.

For GCC’s status to be “useful as a toy for demos”, it’s probably a disservice to even have it in the default builds of GCC until someone is willing to fund the couple of person-years that Jim mentions to get it on track. At least bugreports to LLVM are likely to get traction as it’s actively developed.

For now, if you’re developing vector code on RISC-V, prepare to pair your toolchain with the chip/simulator you’re using. It’s likely to be finicky for a while.

Leave a Reply

Your email address will not be published.

You may use these HTML tags and attributes:

<a href="" title=""> <abbr title=""> <acronym title=""> <b> <blockquote cite=""> <cite> <code> <del datetime=""> <em> <i> <q cite=""> <s> <strike> <strong>