unofficial mirror of libc-alpha@sourceware.org
 help / color / mirror / Atom feed
From: Adhemerval Zanella Netto <adhemerval.zanella@linaro.org>
To: "Christoph Müllner" <christoph.muellner@vrull.eu>
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: Wed, 24 Apr 2024 10:36:43 -0300	[thread overview]
Message-ID: <51b25f25-34a5-4dd6-9061-70bf706724ed@linaro.org> (raw)
In-Reply-To: <CAEg0e7iGC-wyOV+siGvjTYAoFoVpCGncX5UbZN_ep=wfHxbPVw@mail.gmail.com>



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 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.

> 
>>> +#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:[~2024-04-24 13:37 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 [this message]
2024-04-26 11:40         ` Christoph Müllner
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=51b25f25-34a5-4dd6-9061-70bf706724ed@linaro.org \
    --to=adhemerval.zanella@linaro.org \
    --cc=andrew@sifive.com \
    --cc=christoph.muellner@vrull.eu \
    --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).