Re: [PATCH v2 0/4] remove driver-specific logic for AVX builds

[Bruce Richardson <bruce.richardson@intel.com>]

On Mon, Mar 17, 2025 at 10:50:24AM +0100, David Marchand wrote:
> On Fri, Mar 14, 2025 at 6:44 PM Bruce Richardson
> <bruce.richardson@intel.com> wrote:
> >
> > A number of drivers had special optimized AVX2 and AVX512 code paths
> > for performance reasons, and these tended to have copy-pasted logic
> > to build those files. Centralise that logic in the main
> > drivers/meson.build file to avoid duplication.
> >
> > v2: add patch 4 to remove use of unnecessary CC_AVX2_SUPPORT flag
> >
> > Bruce Richardson (4):
> >   build: add generalized AVX handling for drivers
> >   net/intel: use common AVX build code
> >   drivers/net: build use common AVX handling
> >   drivers/net: remove AVX2 build-time define
> >
> >  drivers/meson.build                   | 30 +++++++++++++++++++++++++++
> >  drivers/net/bnxt/bnxt_ethdev.c        |  2 --
> >  drivers/net/bnxt/meson.build          | 10 +--------
> >  drivers/net/enic/meson.build          | 10 +--------
> >  drivers/net/intel/i40e/meson.build    | 26 ++---------------------
> >  drivers/net/intel/iavf/meson.build    | 25 ++--------------------
> >  drivers/net/intel/ice/meson.build     | 25 ++--------------------
> >  drivers/net/intel/idpf/meson.build    | 25 ++--------------------
> >  drivers/net/nfp/meson.build           | 10 +--------
> >  drivers/net/octeon_ep/meson.build     | 13 +-----------
> >  drivers/net/octeon_ep/otx_ep_ethdev.c |  4 ----
> >  drivers/net/virtio/meson.build        |  9 +-------
> >  12 files changed, 43 insertions(+), 146 deletions(-)
> 
> event/dlb2 can be converted too, or is there something special about
> this driver?
> 

Ideally, yes, but it's also a bit special in that it only builds *either*
the SSE code path or the AVX-512 one, and has duplicate symbols/functions
in the two files. This is a bit strange, and probably not what we want, so
I may need to do some work on it before converting it to this scheme.

> 
> As part of the discussion on the base drivers, and seeing this series,
> it would be cool if we had a generic framework to specify a set of
> cflags for a set of sources.

I'd rather not have that. Meson doesn't support having per-c-file flags,
and I don't think we should resort to massive amounts of hackery to try and
support this. I think having the same set of cflags across a whole
component is a good principle, one which we should only violate when we
have to e.g. for base dirs, and for ISA-specific files.

> But at least this series lgtm and is a first step.
> 
Thanks.
/Bruce