From: Brad Smith <brad@comstyle.com>
Subject: Re: UPDATE: highway - riscv64
To: ports@openbsd.org
Date: Fri, 9 Oct 2026 15:24:29 -0400

On 2026-10-09 7:42 a.m., Jeremie Courreges-Anglas wrote:
> On Thu, Oct 08, 2026 at 12:40:13AM +0200, Jeremie Courreges-Anglas wrote:
>> On Fri, Sep 25, 2026 at 08:39:11PM -0400, Brad Smith wrote:
> [...]
>>> -# Don't assume risc-v Vector support
>>> -CONFIGURE_ARGS+=	-DHWY_CMAKE_RVV=OFF
>> This has to stay, we don't want to assume that the vector feature is
>> available.
> ... so I have kept these lines in what I have just committed.


Yes, you're right about this part. I looked into why this exists and its a
workaround for two (older) compiler bugs. With Clang (older than 20) the
Vector header used to only expose Vector builtins if the option
-march=rv64gcv1p0 was used which was removed. With GCC (older
than 15.2) there was an issue with Vector builtins and an interaction
with LTO which resulted in them being disabled unless -march=rv64gcv1p0
was used as a workaround.

I'll see about possibly getting this removed now that the root issues
have been fixed.

>> I had another diff that was more complete.
> Actually I was confused, I had local patches to enable runtime
> detection through elf_aux_info(3) and I had not noticed your latest
> patches committed by Landry in July.
>