Re: [PATCH v4 1/2] arm: lib: xor-neon: remove unnecessary GCC < 4.6 warning

[Arvind Sankar <nivedita@alum.mit.edu>]

On Wed, Jan 20, 2021 at 03:09:53PM -0800, Nick Desaulniers wrote:
> On Tue, Jan 19, 2021 at 1:35 PM Arnd Bergmann <arnd@kernel.org> wrote:
> >
> > On Tue, Jan 19, 2021 at 10:18 PM 'Nick Desaulniers' via Clang Built
> > Linux <clang-built-linux@googlegroups.com> wrote:
> > >
> > > On Tue, Jan 19, 2021 at 5:17 AM Adrian Ratiu <adrian.ratiu@collabora.com> wrote:
> > > > diff --git a/arch/arm/lib/xor-neon.c b/arch/arm/lib/xor-neon.c
> > > > index b99dd8e1c93f..f9f3601cc2d1 100644
> > > > --- a/arch/arm/lib/xor-neon.c
> > > > +++ b/arch/arm/lib/xor-neon.c
> > > > @@ -14,20 +14,22 @@ MODULE_LICENSE("GPL");
> > > >  #error You should compile this file with '-march=armv7-a -mfloat-abi=softfp -mfpu=neon'
> > > >  #endif
> > > >
> > > > +/*
> > > > + * TODO: Even though -ftree-vectorize is enabled by default in Clang, the
> > > > + * compiler does not produce vectorized code due to its cost model.
> > > > + * See: https://github.com/ClangBuiltLinux/linux/issues/503
> > > > + */
> > > > +#ifdef CONFIG_CC_IS_CLANG
> > > > +#warning Clang does not vectorize code in this file.
> > > > +#endif
> > >
> > > Arnd, remind me again why it's a bug that the compiler's cost model
> > > says it's faster to not produce a vectorized version of these loops?
> > > I stand by my previous comment: https://bugs.llvm.org/show_bug.cgi?id=40976#c8
> >
> > The point is that without vectorizing the code, there is no point in building
> > both the default xor code and a "neon" version that has to save/restore
> > the neon registers but doesn't actually use them.
> 
> Doesn't that already happen today with GCC when the pointer arguments
> are overlapping?  The loop is "versioned" and may not actually use the
> NEON registers at all at runtime for such arguments.
> https://godbolt.org/z/q48q8v  See also:
> https://bugs.llvm.org/show_bug.cgi?id=40976#c11.  Or am I missing
> something?

The gcc version is at least useful when the arguments _don't_ overlap,
which is presumably most/all the time.

There's no point to building this file if it's not going to produce a
vectorized version at all. The warning seems unnecessary, and it's not
really a compiler bug either -- until we agree on a way to get clang to
produce a vectorized version, the best thing would be for the neon
version to not be built at all with clang. Is that too messy to achieve?

> 
> So I'm thinking if we extend out this pattern to the rest of the
> functions, we can actually avoid calls to
> kernel_neon_begin()/kernel_neon_end() for cases in which pointers
> would be too close to use the vectorized loop version; meaning for GCC
> this would be an optimization (don't save neon registers when you're
> not going to use them).  I would probably consider moving
> include/asm-generic/xor.h somewhere under arch/arm/
> perhaps...err...something for the other users of <asm-generic/xor.h>.

We can't directly do the patch below since there are other users of the
asm-generic/xor.h implementations than just the neon file.  If it's too
much work to check and add __restrict everywhere, I think we'd either
need to copy the code into the xor-neon file, or maybe do some ifdeffery
so __restrict is only used for the neon version.

> 
> diff --git a/arch/arm/include/asm/xor.h b/arch/arm/include/asm/xor.h
> index aefddec79286..abd748d317e8 100644
> --- a/arch/arm/include/asm/xor.h
> +++ b/arch/arm/include/asm/xor.h
> @@ -148,12 +148,12 @@ extern struct xor_block_template const
> xor_block_neon_inner;
>  static void
>  xor_neon_2(unsigned long bytes, unsigned long *p1, unsigned long *p2)
>  {
> -       if (in_interrupt()) {
> -               xor_arm4regs_2(bytes, p1, p2);
> -       } else {
> +       if (!in_interrupt() && abs((uintptr_t)p1, (uintptr_t)p2) >= 8) {
>                 kernel_neon_begin();
>                 xor_block_neon_inner.do_2(bytes, p1, p2);
>                 kernel_neon_end();
> +       } else {
> +               xor_arm4regs_2(bytes, p1, p2);
>         }
>  }
> diff --git a/arch/arm/lib/xor-neon.c b/arch/arm/lib/xor-neon.c
> index b99dd8e1c93f..0e8e474c0523 100644
> --- a/arch/arm/lib/xor-neon.c
> +++ b/arch/arm/lib/xor-neon.c
> @@ -14,22 +14,6 @@ MODULE_LICENSE("GPL");
>  #error You should compile this file with '-march=armv7-a
> -mfloat-abi=softfp -mfpu=neon'
>  #endif
> 
> -/*
> - * Pull in the reference implementations while instructing GCC (through
> - * -ftree-vectorize) to attempt to exploit implicit parallelism and emit
> - * NEON instructions.
> - */
> -#if __GNUC__ > 4 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 6)
> -#pragma GCC optimize "tree-vectorize"
> -#else
> -/*
> - * While older versions of GCC do not generate incorrect code, they fail to
> - * recognize the parallel nature of these functions, and emit plain ARM code,
> - * which is known to be slower than the optimized ARM code in asm-arm/xor.h.
> - */
> -#warning This code requires at least version 4.6 of GCC
> -#endif
> -
>  #pragma GCC diagnostic ignored "-Wunused-variable"
>  #include <asm-generic/xor.h>
> diff --git a/include/asm-generic/xor.h b/include/asm-generic/xor.h
> index b62a2a56a4d4..69df62095c33 100644
> --- a/include/asm-generic/xor.h
> +++ b/include/asm-generic/xor.h
> @@ -8,7 +8,7 @@
>  #include <linux/prefetch.h>
> 
>  static void
> -xor_8regs_2(unsigned long bytes, unsigned long *p1, unsigned long *p2)
> +xor_8regs_2(unsigned long bytes, unsigned long * __restrict p1,
> unsigned long * __restrict p2)
>  {
>         long lines = bytes / (sizeof (long)) / 8;
> 
> 
> -- 
> Thanks,
> ~Nick Desaulniers