Re: [PATCH v4 3/3] riscv: Add and use alignment-ignorant memcpy

[Evan Green <evan@rivosinc.com>]

On Fri, Jul 7, 2023 at 7:17 PM Stefan O'Rear <sorear@fastmail.com> wrote:
>
>
>
> On Thu, Jul 6, 2023, at 3:29 PM, Evan Green wrote:
> > For CPU implementations that can perform unaligned accesses with little
> > or no performance penalty, create a memcpy implementation that does not
> > bother aligning buffers. It will use a block of integer registers, a
> > single integer register, and fall back to bytewise copy for the
> > remainder.
> >
> > Signed-off-by: Evan Green <evan@rivosinc.com>
> > Reviewed-by: Palmer Dabbelt <palmer@rivosinc.com>
> >
> > ---
> >
> > Changes in v4:
> >  - Fixed comment style (Florian)
> >
> > Changes in v3:
> >  - Word align dest for large memcpy()s.
> >  - Add tags
> >  - Remove spurious blank line from sysdeps/riscv/memcpy.c
> >
> > Changes in v2:
> >  - Used _MASK instead of _FAST value itself.
> >
> >
> > ---
> >  sysdeps/riscv/memcopy.h                       |  26 ++++
> >  sysdeps/riscv/memcpy.c                        |  64 +++++++++
> >  sysdeps/riscv/memcpy_noalignment.S            | 121 ++++++++++++++++++
> >  sysdeps/unix/sysv/linux/riscv/Makefile        |   4 +
> >  .../unix/sysv/linux/riscv/memcpy-generic.c    |  24 ++++
> >  5 files changed, 239 insertions(+)
> >  create mode 100644 sysdeps/riscv/memcopy.h
> >  create mode 100644 sysdeps/riscv/memcpy.c
> >  create mode 100644 sysdeps/riscv/memcpy_noalignment.S
> >  create mode 100644 sysdeps/unix/sysv/linux/riscv/memcpy-generic.c
> >
> > diff --git a/sysdeps/riscv/memcopy.h b/sysdeps/riscv/memcopy.h
> > new file mode 100644
> > index 0000000000..2b685c8aa0
> > --- /dev/null
> > +++ b/sysdeps/riscv/memcopy.h
> > @@ -0,0 +1,26 @@
> > +/* memcopy.h -- definitions for memory copy functions. RISC-V version.
> > +   Copyright (C) 2023 Free Software Foundation, Inc.
> > +   This file is part of the GNU C Library.
> > +
> > +   The GNU C Library is free software; you can redistribute it and/or
> > +   modify it under the terms of the GNU Lesser General Public
> > +   License as published by the Free Software Foundation; either
> > +   version 2.1 of the License, or (at your option) any later version.
> > +
> > +   The GNU C Library is distributed in the hope that it will be useful,
> > +   but WITHOUT ANY WARRANTY; without even the implied warranty of
> > +   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
> > +   Lesser General Public License for more details.
> > +
> > +   You should have received a copy of the GNU Lesser General Public
> > +   License along with the GNU C Library; if not, see
> > +   <https://www.gnu.org/licenses/>.  */
> > +
> > +#include <sysdeps/generic/memcopy.h>
> > +
> > +/* Redefine the generic memcpy implementation to __memcpy_generic, so
> > +   the memcpy ifunc can select between generic and special versions.
> > +   In rtld, don't bother with all the ifunciness. */
> > +#if IS_IN (libc)
> > +#define MEMCPY __memcpy_generic
> > +#endif
> > diff --git a/sysdeps/riscv/memcpy.c b/sysdeps/riscv/memcpy.c
> > new file mode 100644
> > index 0000000000..fdb8dc3208
> > --- /dev/null
> > +++ b/sysdeps/riscv/memcpy.c
> > @@ -0,0 +1,64 @@
> > +/* Multiple versions of memcpy.
> > +   All versions must be listed in ifunc-impl-list.c.
> > +   Copyright (C) 2017-2023 Free Software Foundation, Inc.
> > +   This file is part of the GNU C Library.
> > +
> > +   The GNU C Library is free software; you can redistribute it and/or
> > +   modify it under the terms of the GNU Lesser General Public
> > +   License as published by the Free Software Foundation; either
> > +   version 2.1 of the License, or (at your option) any later version.
> > +
> > +   The GNU C Library is distributed in the hope that it will be useful,
> > +   but WITHOUT ANY WARRANTY; without even the implied warranty of
> > +   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
> > +   Lesser General Public License for more details.
> > +
> > +   You should have received a copy of the GNU Lesser General Public
> > +   License along with the GNU C Library; if not, see
> > +   <https://www.gnu.org/licenses/>.  */
> > +
> > +#if IS_IN (libc)
> > +/* Redefine memcpy so that the compiler won't complain about the type
> > +   mismatch with the IFUNC selector in strong_alias, below.  */
> > +# undef memcpy
> > +# define memcpy __redirect_memcpy
> > +# include <string.h>
> > +#include <ifunc-init.h>
> > +#include <sys/hwprobe.h>
> > +
> > +#define INIT_ARCH()
> > +
> > +extern __typeof (__redirect_memcpy) __libc_memcpy;
> > +
> > +extern __typeof (__redirect_memcpy) __memcpy_generic attribute_hidden;
> > +extern __typeof (__redirect_memcpy) __memcpy_noalignment
> > attribute_hidden;
> > +
> > +static inline __typeof (__redirect_memcpy) *
> > +select_memcpy_ifunc (void)
> > +{
> > +  INIT_ARCH ();
> > +
> > +  struct riscv_hwprobe pair;
> > +
> > +  pair.key = RISCV_HWPROBE_KEY_CPUPERF_0;
> > +  if (__riscv_hwprobe(&pair, 1, 0, NULL, 0) != 0)
> > +    return __memcpy_generic;
> > +
> > +  if ((pair.key > 0) &&
> > +      (pair.value & RISCV_HWPROBE_MISALIGNED_MASK) ==
> > +       RISCV_HWPROBE_MISALIGNED_FAST)
> > +    return __memcpy_noalignment;
>
> It's unclear whether this is semantically correct as a use of
> __riscv_hwprobe.  [1] describes the result of hwprobe as "what's possible
> to enable", leaving open the possibility that additional system calls are

Right, think of it like "cpuid for risc-v".

> needed to determine whether unaligned accesses are supported right now in
> the current process, and [2] adds an (inherited, IIUC) prctl for
> unaligned access which doesn't affect the return value of hwprobe and
> would break this code as written.
>
> (There is nothing in either the privileged spec or the SBI spec to
> prohibit an implementation which provides FAST unaligned access from
> supporting an optional strict alignment checking mode and making it
> available through fw_feature.)

I think your point about prctls() muddying the waters of what hwprobe
is trying to report is generally valid. However in this case, the
prctl() proposed in [2] only disables unaligned accesses for systems
where Linux is handling the trapping/emulation of unaligned accesses.
These systems would only ever report SLOW/EMULATED out of hwprobe.
Since we're only making changes on FAST systems, the prctl()'s setting
doesn't come into play.

-Evan

>
> -s
>
> [1]: https://lore.kernel.org/linux-riscv/mhng-97928779-5d76-4390-a84c-398fdc6a0a4f@palmer-ri-x1c9/
> [2]: https://lore.kernel.org/linux-riscv/20230624122049.7886-6-cleger@rivosinc.com/
>