Re: [PATCH v2] riscv, bpf: Optimize zextw insn with Zba extension

From: Conor Dooley
Date: Wed May 15 2024 - 05:33:38 EST


On Wed, May 15, 2024 at 09:19:46AM +0100, Conor Dooley wrote:
> On Tue, May 14, 2024 at 03:37:02PM +0200, Andrew Jones wrote:
> > On Tue, May 14, 2024 at 07:36:04AM GMT, Wang, Xiao W wrote:
> > > > From: Andrew Jones <ajones@xxxxxxxxxxxxxxxx>
> >> > > > +config RISCV_ISA_ZBA
> > > > > + bool "Zba extension support for bit manipulation instructions"
> > > > > + depends on TOOLCHAIN_HAS_ZBA
> > > >
> > > > We handcraft the instruction, so why do we need toolchain support?
> > >
> > > Good point, we don't need toolchain support for this bpf jit case.
> > >
> > > >
> > > > > + depends on RISCV_ALTERNATIVE
> > > >
> > > > Also, while riscv_has_extension_likely() will be accelerated with
> > > > RISCV_ALTERNATIVE, it's not required.
> > >
> > > Agree, it's not required. For this bpf jit case, we should drop these two dependencies.
> > >
> > > BTW, Zbb is used in bpf jit, the usage there also doesn't depend on toolchain and
> > > RISCV_ALTERNATIVE, but the Kconfig for RISCV_ISA_ZBB has forced the dependencies
> > > due to Zbb assembly programming elsewhere.
> > > Maybe we could just dynamically check the existence of RISCV_ISA_ZB* before jit code
> > > emission? or introduce new config options for bpf jit? I prefer the first method and
> > > welcome any comments.
> >
> > My preferences is to remove as much of the TOOLCHAIN_HAS_ stuff as
> > possible. We should audit the extensions which have them to see if
> > they're really necessary.
>
> While I think it is reasonable to allow the "RISCV_ISA_ZBB" option to
> control whether or not bpf is allowed to use it for optimisations, only
> allowing bpf to do that if there's toolchain support feels odd to me..
> Maybe we need to sorta steal from Charlie's patchset and introduce
> some hidden options that have the toolchain dep that are used by the
> alternative macros etc?
>
> I'll have a poke at how bad that looks I think.

I don't love this, in particular my option naming, but it would allow
the Zbb optimisations in the kernel to not depend on toolchain support
while not muddying the Kconfig waters for users:
https://git.kernel.org/pub/scm/linux/kernel/git/conor/linux.git/commit/?h=riscv-zbb_split
A similar model could be followed if there were to be some
optimisations for Zba in the future that do require toolchain support:

Attachment: signature.asc
Description: PGP signature