Re: [PATCH 2/2] arm: lib: xor-neon: disable clang vectorization

From: Nick Desaulniers <ndesaulniers@google.com>
To: Arvind Sankar <nivedita@alum.mit.edu>
Cc: Adrian Ratiu <adrian.ratiu@collabora.com>,
	Nathan Chancellor <natechancellor@gmail.com>,
	Arnd Bergmann <arnd@arndb.de>,
	Linux ARM <linux-arm-kernel@lists.infradead.org>,
	clang-built-linux <clang-built-linux@googlegroups.com>,
	Russell King <linux@armlinux.org.uk>,
	LKML <linux-kernel@vger.kernel.org>,
	Collabora Kernel ML <kernel@collabora.com>,
	Ard Biesheuvel <ardb@kernel.org>
Subject: Re: [PATCH 2/2] arm: lib: xor-neon: disable clang vectorization
Date: Tue, 10 Nov 2020 14:36:27 -0800	[thread overview]
Message-ID: <CAKwvOd=NQ==umC+N_Sgji5HCCFTRARh4jWiB3DaBfV6jDd5cRg@mail.gmail.com> (raw)
In-Reply-To: <20201110221511.GA1373528@rani.riverdale.lan>

On Tue, Nov 10, 2020 at 2:15 PM Arvind Sankar <nivedita@alum.mit.edu> wrote:
>
> On Tue, Nov 10, 2020 at 01:41:17PM -0800, Nick Desaulniers wrote:
> > On Mon, Nov 9, 2020 at 11:51 AM Adrian Ratiu <adrian.ratiu@collabora.com> wrote:
> > >
> > > On Fri, 06 Nov 2020, Nick Desaulniers <ndesaulniers@google.com>
> > > wrote:
> > > > +#pragma clang loop vectorize(enable)
> > > >         do {
> > > >                 p1[0] ^= p2[0] ^ p3[0] ^ p4[0] ^ p5[0]; p1[1] ^=
> > > >                 p2[1] ^ p3[1] ^ p4[1] ^ p5[1];
> > > > ``` seems to generate the vectorized code.
> > > >
> > > > Why don't we find a way to make those pragma's more toolchain
> > > > portable, rather than open coding them like I have above rather
> > > > than this series?
> > >
> > > Hi again Nick,
> > >
> > > How did you verify the above pragmas generate correct vectorized
> > > code?  Have you tested this specific use case?
> >
> > I read the disassembly before and after my suggested use of pragmas;
> > look for vld/vstr.  You can also add -Rpass-missed=loop-vectorize to
> > CFLAGS_xor-neon.o in arch/arm/lib/Makefile and rebuild
> > arch/arm/lib/xor-neon.o with CONFIG_BTRFS enabled.
> >
>
> https://godbolt.org/z/1oo9M6
>
> With the __restrict__ keywords added, clang seems to vectorize the loop,
> but still reports that vectorization wasn't beneficial -- any idea
> what's going on?

I suspect that loop-vectorize is a higher level pass that relies on
slp-vectorizer for the transform.

$ clang -O2 --target=arm-linux-gnueabi -S -o - foo.c -mfpu=neon -mllvm
-print-after-all
...
*** IR Dump After SLP Vectorizer ***
(bunch of <4 x i32> types)

If you add -Rpass-missed=slp-vectorizer, observe that the existing
warnings from -Rpass-missed=loop-vectorize disappear; I suspect
loop-vectorize will print a "remark" if passes it calls did not, but
returned some for of error code.

-Rpass=slp-vectorizer shows that it vectorizes two sequences of the
loop, and warns that some third portion (that's
non-immediately-obvious to me) was non beneficial.

-- 
Thanks,
~Nick Desaulniers