unofficial mirror of libc-alpha@sourceware.org
 help / color / mirror / Atom feed
From: "Christoph Müllner" <christoph.muellner@vrull.eu>
To: Adhemerval Zanella Netto <adhemerval.zanella@linaro.org>
Cc: libc-alpha@sourceware.org, Palmer Dabbelt <palmer@dabbelt.com>,
	 Darius Rad <darius@bluespec.com>,
	Andrew Waterman <andrew@sifive.com>,
	 Philipp Tomsich <philipp.tomsich@vrull.eu>,
	Evan Green <evan@rivosinc.com>, DJ Delorie <dj@redhat.com>,
	 Vineet Gupta <vineetg@rivosinc.com>,
	Kito Cheng <kito.cheng@sifive.com>,
	 Jeff Law <jeffreyalaw@gmail.com>
Subject: Re: [PATCH 2/7] RISC-V: Add Zbb optimized memchr as ifunc
Date: Fri, 26 Apr 2024 13:40:59 +0200	[thread overview]
Message-ID: <CAEg0e7jAkbbpYS_2CXVFgZNQ6in_v2EbN_HvO_PdzCW8DtJU-w@mail.gmail.com> (raw)
In-Reply-To: <51b25f25-34a5-4dd6-9061-70bf706724ed@linaro.org>

On Wed, Apr 24, 2024 at 3:36 PM Adhemerval Zanella Netto
<adhemerval.zanella@linaro.org> wrote:
>
>
>
> On 24/04/24 10:16, Christoph Müllner wrote:
> > On Wed, Apr 24, 2024 at 2:53 PM Adhemerval Zanella Netto
> > <adhemerval.zanella@linaro.org> wrote:
> >>
> >>
> >>
> >> On 22/04/24 04:43, Christoph Müllner wrote:
> >>> When building with Zbb enabled, memchr benefits from using orc.b in
> >>> find_zero_all().  This patch changes the build system such, that a
> >>> non-Zbb version as well as a Zbb version of this routine is built.
> >>> Further, a ifunc resolver is provided that selects the right routine
> >>> based on the outcome of extension probing via hwprobe().
> >>>
> >>> Signed-off-by: Christoph Müllner <christoph.muellner@vrull.eu>
> >>> ---
> >>>  sysdeps/riscv/multiarch/memchr-generic.c      | 26 +++++++++
> >>>  sysdeps/riscv/multiarch/memchr-zbb.c          | 30 ++++++++++
> >>>  .../unix/sysv/linux/riscv/multiarch/Makefile  |  3 +
> >>>  .../linux/riscv/multiarch/ifunc-impl-list.c   | 31 ++++++++--
> >>>  .../unix/sysv/linux/riscv/multiarch/memchr.c  | 57 +++++++++++++++++++
> >>>  5 files changed, 142 insertions(+), 5 deletions(-)
> >>>  create mode 100644 sysdeps/riscv/multiarch/memchr-generic.c
> >>>  create mode 100644 sysdeps/riscv/multiarch/memchr-zbb.c
> >>>  create mode 100644 sysdeps/unix/sysv/linux/riscv/multiarch/memchr.c
> >>>
> >>> diff --git a/sysdeps/riscv/multiarch/memchr-generic.c b/sysdeps/riscv/multiarch/memchr-generic.c
> >>> new file mode 100644
> >>> index 0000000000..a96c36398b
> >>> --- /dev/null
> >>> +++ b/sysdeps/riscv/multiarch/memchr-generic.c
> >>> @@ -0,0 +1,26 @@
> >>> +/* Re-include the default memchr implementation.
> >>> +   Copyright (C) 2024 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 <string.h>
> >>> +
> >>> +#if IS_IN(libc)
> >>> +# define MEMCHR __memchr_generic
> >>> +# undef libc_hidden_builtin_def
> >>> +# define libc_hidden_builtin_def(x)
> >>> +#endif
> >>> +#include <string/memchr.c>
> >>> diff --git a/sysdeps/riscv/multiarch/memchr-zbb.c b/sysdeps/riscv/multiarch/memchr-zbb.c
> >>> new file mode 100644
> >>> index 0000000000..bead0335ae
> >>> --- /dev/null
> >>> +++ b/sysdeps/riscv/multiarch/memchr-zbb.c
> >>> @@ -0,0 +1,30 @@
> >>> +/* Re-include the default memchr implementation for Zbb.
> >>> +   Copyright (C) 2024 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 <string.h>
> >>> +
> >>> +#if IS_IN(libc)
> >>> +# define MEMCHR __memchr_zbb
> >>> +# undef libc_hidden_builtin_def
> >>> +# define libc_hidden_builtin_def(x)
> >>> +#endif
> >>> +/* Convince preprocessor to have Zbb instructions.  */
> >>> +#ifndef __riscv_zbb
> >>> +# define __riscv_zbb
> >>> +#endif
> >>
> >> Is there a way to specific the compiler to enable a extension, like aarch64
> >> -march=arch{+[no]feature}? I think ideally this should be enabled as CFLAGS
> >> instead of messing with compiler defined pre-processor.
> >
> > The tools expect a list of all extensions as parameter to the -march= option.
> > But there is no way to append extensions to an existing march string
> > on the command line.
> >
> > And if we would add this feature today, it would take many years until we could
> > use it here, because we want to remain compatible with old tools.
> > Or we enable the optimization only when being built with new tools, but that
> > adds even more complexity and build/test configurations.
> >
> > What we have is:
> > * Preprocessor (since forever): Extension test macros (__riscv_EXTENSION)
> > * Command line (since forever): -march=BASE_EXTENSIONLIST
> > * GAS (since Nov 21): .option arch, +EXTENSION (in combination with
> > option push/pop)
> > * GCC (since Nov 23): __attribute__((target("arch=+EXTENSION")))
> >
> > I was not sure about using __riscv_zbb as well, but I considered it safe within
> > ifdef tests that ensure the macro won't be set twice.
> > If that's a concern, I could change to use something like this:
> > #define __riscv_force_zbb
> > #include <impl.c>
> > #undef __riscv_force_zbb
> > ... and change string-fza.h like this:
> > #if defined(__riscv_zbb) || defined(__riscv_force_zbb)
> > // orc.b
> > #endif
> >
> > BR
> > Christoph
>
> Another options would to parse the current march and add the extension if required,
> something like:
>
> abi=$(riscv64-linux-gnu-gcc -Q --help=target | grep march | cut -d '=' -f2 | xargs)
> if [[ ! "$abi" =~ "_zbb" ]]
> then
>   abi="$abi"_zbb
> fi

I tried to work something out, but this attempt also won't work reliably.
Until recently (Jan 5, 2024) GCC required that the extensions on the command
line were provided in canonical order. I.e., "_zbb" would probably have to be
inserted somewhere in the middle. Sorting the tokens of the march string seems
to be an overkill.

> I don't have a strong preference, it is just that by not using the compiler flag
> we won't be able to either use the builtin (__builtin_riscv_orc_b_32) and/or get
> a possible better code generation from compiler.

I agree that builtins are better.
At the same time, the solution in the patch is no worse than what we
already have
in glibc right now. Further, this builtin was added to GCC on Jan 15, 2024.
So, it can only be used if the build tools are recent enough.

>
> >
> >>> +#include <string/memchr.c>
> >>> diff --git a/sysdeps/unix/sysv/linux/riscv/multiarch/Makefile b/sysdeps/unix/sysv/linux/riscv/multiarch/Makefile
> >>> index fcef5659d4..5586d11c89 100644
> >>> --- a/sysdeps/unix/sysv/linux/riscv/multiarch/Makefile
> >>> +++ b/sysdeps/unix/sysv/linux/riscv/multiarch/Makefile
> >>> @@ -1,5 +1,8 @@
> >>>  ifeq ($(subdir),string)
> >>>  sysdep_routines += \
> >>> +  memchr \
> >>> +  memchr-generic \
> >>> +  memchr-zbb \
> >>>    memcpy \
> >>>    memcpy-generic \
> >>>    memcpy_noalignment \
> >>> diff --git a/sysdeps/unix/sysv/linux/riscv/multiarch/ifunc-impl-list.c b/sysdeps/unix/sysv/linux/riscv/multiarch/ifunc-impl-list.c
> >>> index 9f806d7a9e..7321144a32 100644
> >>> --- a/sysdeps/unix/sysv/linux/riscv/multiarch/ifunc-impl-list.c
> >>> +++ b/sysdeps/unix/sysv/linux/riscv/multiarch/ifunc-impl-list.c
> >>> @@ -20,19 +20,40 @@
> >>>  #include <string.h>
> >>>  #include <sys/hwprobe.h>
> >>>
> >>> +#define ARRAY_SIZE(A) (sizeof (A) / sizeof ((A)[0]))
> >>> +
> >>>  size_t
> >>>  __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array,
> >>>                       size_t max)
> >>>  {
> >>>    size_t i = max;
> >>> +  struct riscv_hwprobe pairs[] = {
> >>> +    { .key = RISCV_HWPROBE_KEY_IMA_EXT_0 },
> >>> +    { .key = RISCV_HWPROBE_KEY_CPUPERF_0 },
> >>> +  };
> >>>
> >>> +  bool has_zbb = false;
> >>>    bool fast_unaligned = false;
> >>>
> >>> -  struct riscv_hwprobe pair = { .key = RISCV_HWPROBE_KEY_CPUPERF_0 };
> >>> -  if (__riscv_hwprobe (&pair, 1, 0, NULL, 0) == 0
> >>> -      && (pair.value & RISCV_HWPROBE_MISALIGNED_MASK)
> >>> -          == RISCV_HWPROBE_MISALIGNED_FAST)
> >>> -    fast_unaligned = true;
> >>> +  if (__riscv_hwprobe (pairs, ARRAY_SIZE (pairs), 0, NULL, 0) == 0)
> >>> +    {
> >>> +      struct riscv_hwprobe *pair;
> >>> +
> >>> +      /* RISCV_HWPROBE_KEY_IMA_EXT_0  */
> >>> +      pair = &pairs[0];
> >>> +      if (pair->value & RISCV_HWPROBE_EXT_ZBB)
> >>> +        has_zbb = true;
> >>> +
> >>> +      /* RISCV_HWPROBE_KEY_CPUPERF_0  */
> >>> +      pair = &pairs[1];
> >>> +      if ((pair->value & RISCV_HWPROBE_MISALIGNED_MASK)
> >>> +        == RISCV_HWPROBE_MISALIGNED_FAST)
> >>> +        fast_unaligned = true;
> >>> +    }
> >>> +
> >>> +  IFUNC_IMPL (i, name, memchr,
> >>> +           IFUNC_IMPL_ADD (array, i, memchr, has_zbb, __memchr_zbb)
> >>> +           IFUNC_IMPL_ADD (array, i, memchr, 1, __memchr_generic))
> >>>
> >>>    IFUNC_IMPL (i, name, memcpy,
> >>>             IFUNC_IMPL_ADD (array, i, memcpy, fast_unaligned,
> >>> diff --git a/sysdeps/unix/sysv/linux/riscv/multiarch/memchr.c b/sysdeps/unix/sysv/linux/riscv/multiarch/memchr.c
> >>> new file mode 100644
> >>> index 0000000000..bc076cbf24
> >>> --- /dev/null
> >>> +++ b/sysdeps/unix/sysv/linux/riscv/multiarch/memchr.c
> >>> @@ -0,0 +1,57 @@
> >>> +/* Multiple versions of memchr.
> >>> +   All versions must be listed in ifunc-impl-list.c.
> >>> +   Copyright (C) 2017-2024 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 memchr so that the compiler won't complain about the type
> >>> +   mismatch with the IFUNC selector in strong_alias, below.  */
> >>> +# undef memchr
> >>> +# define memchr __redirect_memchr
> >>> +# include <stdint.h>
> >>> +# include <string.h>
> >>> +# include <ifunc-init.h>
> >>> +# include <riscv-ifunc.h>
> >>> +# include <sys/hwprobe.h>
> >>> +
> >>> +extern __typeof (__redirect_memchr) __libc_memchr;
> >>> +
> >>> +extern __typeof (__redirect_memchr) __memchr_generic attribute_hidden;
> >>> +extern __typeof (__redirect_memchr) __memchr_zbb attribute_hidden;
> >>> +
> >>> +static inline __typeof (__redirect_memchr) *
> >>> +select_memchr_ifunc (uint64_t dl_hwcap, __riscv_hwprobe_t hwprobe_func)
> >>> +{
> >>> +  unsigned long long int v;
> >>> +  if (__riscv_hwprobe_one (hwprobe_func, RISCV_HWPROBE_KEY_IMA_EXT_0, &v) == 0
> >>> +      && (v & RISCV_HWPROBE_EXT_ZBB))
> >>> +    return __memchr_zbb;
> >>> +
> >>> +  return __memchr_generic;
> >>> +}
> >>> +
> >>> +riscv_libc_ifunc (__libc_memchr, select_memchr_ifunc);
> >>> +
> >>> +# undef memchr
> >>> +strong_alias (__libc_memchr, memchr);
> >>> +# ifdef SHARED
> >>> +__hidden_ver1 (memchr, __GI_memchr, __redirect_memchr)
> >>> +  __attribute__ ((visibility ("hidden"))) __attribute_copy__ (memchr);
> >>> +# endif
> >>> +#else
> >>> +# include <string/memchr.c>
> >>> +#endif

  reply	other threads:[~2024-04-26 11:41 UTC|newest]

Thread overview: 20+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2024-04-22  7:43 [PATCH 0/7] Add ifunc support for existing Zbb optimizations Christoph Müllner
2024-04-22  7:43 ` [PATCH 1/7] RISC-V: Use .insn directive form for orc.b Christoph Müllner
2024-04-22  7:43 ` [PATCH 2/7] RISC-V: Add Zbb optimized memchr as ifunc Christoph Müllner
2024-04-24 12:53   ` Adhemerval Zanella Netto
2024-04-24 13:16     ` Christoph Müllner
2024-04-24 13:36       ` Adhemerval Zanella Netto
2024-04-26 11:40         ` Christoph Müllner [this message]
2024-04-30 15:13         ` Palmer Dabbelt
2024-04-30 17:45           ` Adhemerval Zanella Netto
2024-04-30 17:54             ` Palmer Dabbelt
2024-04-30 18:44               ` Vineet Gupta
2024-05-06 13:20               ` Christoph Müllner
2024-05-06 13:32                 ` Kito Cheng
2024-05-06 13:46                   ` Christoph Müllner
2024-05-06 13:58                     ` Kito Cheng
2024-04-22  7:43 ` [PATCH 3/7] RISC-V: Add Zbb optimized memrchr " Christoph Müllner
2024-04-22  7:44 ` [PATCH 4/7] RISC-V: Add Zbb optimized strchrnul " Christoph Müllner
2024-04-22  7:44 ` [PATCH 5/7] RISC-V: Add Zbb optimized strcmp " Christoph Müllner
2024-04-22  7:44 ` [PATCH 6/7] RISC-V: Add Zbb optimized strlen " Christoph Müllner
2024-04-22  7:44 ` [PATCH 7/7] RISC-V: Add Zbb optimized strncmp " Christoph Müllner

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

  List information: https://www.gnu.org/software/libc/involved.html

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=CAEg0e7jAkbbpYS_2CXVFgZNQ6in_v2EbN_HvO_PdzCW8DtJU-w@mail.gmail.com \
    --to=christoph.muellner@vrull.eu \
    --cc=adhemerval.zanella@linaro.org \
    --cc=andrew@sifive.com \
    --cc=darius@bluespec.com \
    --cc=dj@redhat.com \
    --cc=evan@rivosinc.com \
    --cc=jeffreyalaw@gmail.com \
    --cc=kito.cheng@sifive.com \
    --cc=libc-alpha@sourceware.org \
    --cc=palmer@dabbelt.com \
    --cc=philipp.tomsich@vrull.eu \
    --cc=vineetg@rivosinc.com \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).