llvm dev - Nov 2021 - Proposal: Make the VE target official

Hi all,

NEC have been building up target support for SX-Aurora TSUBASA in LLVM
upstream - the VE target (Vector Engine) [1].  Today VE support in
upstream is sufficient for scalar code generation from C code with
target-specific vector intrinsics.

We propose the switch from 'experimental' backend status to
'official',
throwing into the ring:

* The staging builder (clang-ve-ninja) [2].  This builder builds and
  checks LLVM+Clang and compiler-rt. The functional compiler-rt tests
  run C code on a VE device.
  We will extend this builder to test all other runtime libraries
(libcxxabi, libcxx, ..) for VE as continue upstreaming.
* LIT tests.  VE has extensive lit tests (test/CodeGen/VE) for scalar
  code generation and vector intrinsics.
* D113093 [3] has the changes to make all compiler-rt tests pass on VE.
  Remaining failures are due to denormal support, alignment
  requirements and subtle differences in syscalls.
* The downstream reference implementation (LLVM-VE [4]).  This includes
  compiler-rt,libcxxabi,libcxx and openmp target offloading and will be
  the basis for upstreaming.

What do you say?

- Simon

[1] Vector Engine:
https://www.nec.com/en/global/solutions/hpc/sx/vector_engine.html
[2] clang-ve-ninja (recent run):
https://lab.llvm.org/staging/#/builders/24/builds/6594
[3] compiler-rt patch: https://reviews.llvm.org/D113093
[4] llvm-ve source: https://github.com/sx-aurora-dev/llvm-project.git

On Wed, 3 Nov 2021 at 12:52, Simon Moll <Simon.Moll at emea.nec.com>
wrote:
> We propose the switch from 'experimental' backend status to
'official',
>
Hi Simon,

The VE community seems to have adopted all requirements laid out in the new
targets' document[1], so I guess it's just a matter of ironing out the
transition.

* The staging builder (clang-ve-ninja) [2].  This builder builds
and
>   checks LLVM+Clang and compiler-rt. The functional compiler-rt tests
>   run C code on a VE device.
>
This bot seems red for a long time. One of the ideas that reflect a target
to be stable is that its buildbot is "mostly" green throughout the
warmup
period (and hopefully just before moving out of experimental).

* D113093 [3] has the changes to make all compiler-rt tests pass on
VE.
>   Remaining failures are due to denormal support, alignment
>   requirements and subtle differences in syscalls.
>
Will this make the bots green? It's ok to disable tests that fail for known
reasons (or are irrelevant), but we shouldn't disable tests just to make
the bots green to promote the target.

* The downstream reference implementation (LLVM-VE [4]).  This
includes
>   compiler-rt,libcxxabi,libcxx and openmp target offloading and will be
>   the basis for upstreaming.
>
I looked at the branch and it seems to have a single merged patch that is
not only huge, but has a lot of style changes to unrelated code (ex.
clang-tools-extra).

It would be much simpler if you had a branch with the changes that you want
to upstream into separated patches rebase onto LLVM's main branch.

cheers,
--renato
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20211108/ea499eea/attachment.html>