[v2,7/7] rust-target-config.bbclass: Update for new riscv TUNE_FEATURES

From: Mark Hatle <mark.hatle@amd.com>

Add the new TUNE_FEATURES to the 'features:' list, based on matching output
with:

   rustc --target=riscv32i-unknown-none-elf -Ctarget-feature=help

Use the TUNE_RISCV_ABI instead of guessing for the ABI.

Pass the arch "as-is", since it should now be riscv32 or riscv64.

Signed-off-by: Mark Hatle <mark.hatle@amd.com>
---
 .../classes-recipe/rust-target-config.bbclass | 65 ++++++++++++-------
 meta/lib/oe/rust.py                           |  2 -
 2 files changed, 41 insertions(+), 26 deletions(-)

On Wed, 18 Jun 2025 at 01:39, Mark Hatle via lists.openembedded.org
<mark.hatle=kernel.crashing.org@lists.openembedded.org> wrote:
> -        tspec['llvm-abiname'] = "lp64d"
> +        tspec['llvm-abiname'] = d.getVar('TUNE_RISCV_ABI')
>      if "riscv32" in tspec['llvm-target']:
> -        tspec['llvm-abiname'] = "ilp32d"
> +        tspec['llvm-abiname'] = d.getVar('TUNE_RISCV_ABI')

Unfortunately this adds a dependency on TUNE_RISCV_ABI (which is set
only if the target is a risc-v machine) to rust-native's
rust_gen_targets task, breaking sstate reuse for it between risc-v and
any other target. I can imagine other native pieces written in rust
will have the same issue.

It's not immediately obvious to me how to address this, so I'd really
appreciate suggestions. The code is hairy.

Ticket: https://bugzilla.yoctoproject.org/show_bug.cgi?id=16132

Alex

On Wed, 2026-01-14 at 11:16 +0100, Alexander Kanavin wrote:
> On Wed, 18 Jun 2025 at 01:39, Mark Hatle via lists.openembedded.org
> <mark.hatle=kernel.crashing.org@lists.openembedded.org> wrote:
> > -        tspec['llvm-abiname'] = "lp64d"
> > +        tspec['llvm-abiname'] = d.getVar('TUNE_RISCV_ABI')
> >      if "riscv32" in tspec['llvm-target']:
> > -        tspec['llvm-abiname'] = "ilp32d"
> > +        tspec['llvm-abiname'] = d.getVar('TUNE_RISCV_ABI')
> 
> Unfortunately this adds a dependency on TUNE_RISCV_ABI (which is set
> only if the target is a risc-v machine) to rust-native's
> rust_gen_targets task, breaking sstate reuse for it between risc-v and
> any other target. I can imagine other native pieces written in rust
> will have the same issue.
> 
> It's not immediately obvious to me how to address this, so I'd really
> appreciate suggestions. The code is hairy.
> 
> Ticket: https://bugzilla.yoctoproject.org/show_bug.cgi?id=16132

Nice find. I worry about our sstate tests not covering this. We might need to do:

-rust_gen_target[vardepsexclude] += "ABIEXTENSION llvm_cpu"
+rust_gen_target[vardepsexclude] += "ABIEXTENSION llvm_cpu TUNE_RISCV_ABI"

Cheers,

Richard

On Wed, 14 Jan 2026 at 11:33, Richard Purdie
<richard.purdie@linuxfoundation.org> wrote:

> Nice find. I worry about our sstate tests not covering this. We might need to do:
>
> -rust_gen_target[vardepsexclude] += "ABIEXTENSION llvm_cpu"
> +rust_gen_target[vardepsexclude] += "ABIEXTENSION llvm_cpu TUNE_RISCV_ABI"

I thought of that, but would this rebuild correctly if someone first
builds for one TUNE_RISCV_ABI, then for another? I just don't know
enough about it to run experiments. It probably would, as that
variable feeds into TUNE_CCARGS.

Alex

On Wed, 2026-01-14 at 11:50 +0100, Alexander Kanavin wrote:
> On Wed, 14 Jan 2026 at 11:33, Richard Purdie
> <richard.purdie@linuxfoundation.org> wrote:
> 
> > Nice find. I worry about our sstate tests not covering this. We
> > might need to do:
> > 
> > -rust_gen_target[vardepsexclude] += "ABIEXTENSION llvm_cpu"
> > +rust_gen_target[vardepsexclude] += "ABIEXTENSION llvm_cpu
> > TUNE_RISCV_ABI"
> 
> I thought of that, but would this rebuild correctly if someone first
> builds for one TUNE_RISCV_ABI, then for another? I just don't know
> enough about it to run experiments. It probably would, as that
> variable feeds into TUNE_CCARGS.

The comment above this line says:

# These are accounted for in tmpdir path names so don't need to be in the task sig

so in theory this would handle changes to different machines with
different ABIs fine, as long as those ABIs have different tmpdir paths
which they should.

It doesn't guard against the value of TUNE_RISCV_ABI changing for a
given target as that wouldn't rebuild correctly.

I guess if we wanted to be really clear, we could probably do something
like:

RUSTCONFIG_EXCLUDEVARS = ""
RUSTCONFIG_EXCLUDEVARS:class-native = "ABIEXTENSION llvm_cpu TUNE_RISCV_ABI"
RUSTCONFIG_EXCLUDEVARS:class-nativesdk = "ABIEXTENSION llvm_cpu TUNE_RISCV_ABI"
rust_gen_target[vardepsexclude] += "${RUSTCONFIG_EXCLUDEVARS}"

but I guess we've not thought that was worth the effort in the past...

Cheers,

Richard