On Fri, May 10, 2024 at 11:26:22PM +0100, Conor Dooley wrote:
> On Fri, May 10, 2024 at 02:43:10PM -0700, Charlie Jenkins wrote:
> > On Fri, May 10, 2024 at 09:43:33PM +0100, Conor Dooley wrote:
> > > Hey Charlie,
> > > 
> > > On Tue, May 07, 2024 at 06:36:28PM -0700, Charlie Jenkins wrote:
> > > > Current versions of the kernel add "v" to the march and then immeidately
> > > > filter it out such that "v" is not passed to CFLAGS.  Instead of doing
> > > > this filtering, code blocks in the kernel that want to use vector
> > > > assembly have been changed to locally enable vector (using ".option
> > > > arch, +v").
> > > 
> > > Other content in the series aside, since this is a change that could be
> > > made independently of the main series objectives, I figured it was worth
> > > pointing out that this is not a change without downsides: I think that
> > > it would drop support for vector with most versions of LLVM as
> > > .option arch support there is much more recent thing than it is for gcc.
> > > Off the top of my head I don't know exactly the versions involved, but
> > > it is something like LLVM-14 supports vector but only LLVM-17 and later
> > > supports .option arch.
> > 
> > Toolchain incompatibilities are always fun. It does look like .option
> > arch was introduced in LLVM-17. That would be a regression. We do use
> > .option arch for every other extension, but vector was treated special
> > when it was introduced unfortunately so maybe we have to live with the
> > weird march parsing hack.
> 
> I wrote out a long message about the history of some of this, but right
> at the end I was scrolling through my chat logs with Andy and realised
> we actually did make it depend on AS_HAS_OPTION_ARCH, so it should be
> safe to do without regressing anything. I didn't notice in the diff that
> the AS_HAS_OPTION_ARCH was a movement, not an addition. Maybe Andy knows
> why, despite the dependency on the assembler having it, we didn't use it
> everywhere.

And while I remember, the bits of this doing the .option arch conversion
should be (IMO) split into a different patch to anything dealing with
Kconfig options etc.

Cheers,
Conor.