Re: [PATCH] riscv: use the generic string routines

From: Matteo Croce <mcroce@linux.microsoft.com>
To: David Laight <David.Laight@aculab.com>
Cc: Guo Ren <guoren@kernel.org>, Palmer Dabbelt <palmer@dabbelt.com>,
	linux-riscv <linux-riscv@lists.infradead.org>,
	Linux Kernel Mailing List <linux-kernel@vger.kernel.org>,
	linux-arch <linux-arch@vger.kernel.org>,
	Paul Walmsley <paul.walmsley@sifive.com>,
	Albert Ou <aou@eecs.berkeley.edu>,
	Atish Patra <Atish.Patra@wdc.com>,
	Emil Renner Berthing <kernel@esmil.dk>,
	Akira Tsukamoto <akira.tsukamoto@gmail.com>,
	Drew Fustini <drew@beagleboard.org>,
	Bin Meng <bmeng.cn@gmail.com>,
	Christoph Hellwig <hch@infradead.org>
Subject: Re: [PATCH] riscv: use the generic string routines
Date: Sun, 19 Sep 2021 21:13:24 +0200	[thread overview]
Message-ID: <CAFnufp2M_9_TRxoXbRK0bggPXyTgffYnA4moez=uWDNNb=aT8w@mail.gmail.com> (raw)
In-Reply-To: <9a8137149a164a13a7a04d72b133ad3b@AcuMS.aculab.com>

On Mon, Sep 13, 2021 at 1:35 PM David Laight <David.Laight@aculab.com> wrote:
>
> > > These ended up getting rejected by Linus, so I'm going to hold off on
> > > this for now.  If they're really out of lib/ then I'll take the C
> > > routines in arch/riscv, but either way it's an issue for the next
> > > release.
> > Agree, we should take the C routine in arch/riscv for common
> > implementation. If any vendor what custom implementation they could
> > use the alternative framework in errata for string operations.
>
> I though the asm ones were significantly faster because
> they were less affected by read latency.
>
> (But they were horribly broken for misaligned transfers.)
>

I can get the same exact performance (and a very similar machine code)
in C with this on top of the C memset implementation:

--- a/arch/riscv/lib/string.c
+++ b/arch/riscv/lib/string.c
@@ -112,9 +112,12 @@ EXPORT_SYMBOL(__memmove);
 void *memmove(void *dest, const void *src, size_t count) __weak
__alias(__memmove);
 EXPORT_SYMBOL(memmove);

+#define BATCH 4
+
 void *__memset(void *s, int c, size_t count)
 {
  union types dest = { .as_u8 = s };
+ int i;

  if (count >= MIN_THRESHOLD) {
  unsigned long cu = (unsigned long)c;
@@ -138,8 +141,12 @@ void *__memset(void *s, int c, size_t count)
  }

  /* Copy using the largest size allowed */
- for (; count >= BYTES_LONG; count -= BYTES_LONG)
- *dest.as_ulong++ = cu;
+ for (; count >= BYTES_LONG * BATCH; count -= BYTES_LONG * BATCH) {
+#pragma GCC unroll 4
+     for (i = 0; i < BATCH; i++)
+         dest.as_ulong[i] = cu;
+     dest.as_ulong += BATCH;
+ }
  }

On the BeagleV the memset speed with the different batch size are:

1 (stock): 267 Mb/s
2: 272 Mb/s
4: 276 Mb/s
8: 276 Mb/s

The problem with biggest batch size is that it will fallback to a
single byte copy if the buffers are too small.

Regards,
-- 
per aspera ad upstream
