Re: [PATCH v5 10/17] string: Improve generic memchr

[Noah Goldstein <goldstein.w.n@gmail.com>]

On Mon, Jan 9, 2023 at 12:51 PM Adhemerval Zanella Netto
<adhemerval.zanella@linaro.org> wrote:
>
>
>
> On 05/01/23 20:49, Noah Goldstein wrote:
> > On Mon, Sep 19, 2022 at 1:05 PM Adhemerval Zanella via Libc-alpha
> > <libc-alpha@sourceware.org> wrote:
> >>
> >> New algorithm have the following key differences:
> >>
> >>   - Reads first word unaligned and use string-maskoff function to
> >>     remove unwanted data.  This strategy follow arch-specific
> >>     optimization used on aarch64 and powerpc.
> >>
> >>   - Use string-fz{b,i} and string-opthr functions.
> >>
> >> Checked on x86_64-linux-gnu, i686-linux-gnu, powerpc-linux-gnu,
> >> and powerpc64-linux-gnu by removing the arch-specific assembly
> >> implementation and disabling multi-arch (it covers both LE and BE
> >> for 64 and 32 bits).
> >>
> >> Co-authored-by: Richard Henderson  <rth@twiddle.net>
> >> ---
> >>  string/memchr.c                               | 168 +++++-------------
> >>  .../powerpc32/power4/multiarch/memchr-ppc32.c |  14 +-
> >>  .../powerpc64/multiarch/memchr-ppc64.c        |   9 +-
> >>  3 files changed, 48 insertions(+), 143 deletions(-)
> >>
> >> diff --git a/string/memchr.c b/string/memchr.c
> >> index 422bcd0cd6..08d518b02d 100644
> >> --- a/string/memchr.c
> >> +++ b/string/memchr.c
> >> @@ -1,10 +1,6 @@
> >> -/* Copyright (C) 1991-2022 Free Software Foundation, Inc.
> >> +/* Scan memory for a character.  Generic version
> >> +   Copyright (C) 1991-2022 Free Software Foundation, Inc.
> >>     This file is part of the GNU C Library.
> >> -   Based on strlen implementation by Torbjorn Granlund (tege@sics.se),
> >> -   with help from Dan Sahlin (dan@sics.se) and
> >> -   commentary by Jim Blandy (jimb@ai.mit.edu);
> >> -   adaptation to memchr suggested by Dick Karpinski (dick@cca.ucsf.edu),
> >> -   and implemented by Roland McGrath (roland@ai.mit.edu).
> >>
> >>     The GNU C Library is free software; you can redistribute it and/or
> >>     modify it under the terms of the GNU Lesser General Public
> >> @@ -20,143 +16,65 @@
> >>     License along with the GNU C Library; if not, see
> >>     <https://www.gnu.org/licenses/>.  */
> >>
> >> -#ifndef _LIBC
> >> -# include <config.h>
> >> -#endif
> >> -
> >> +#include <intprops.h>
> >> +#include <string-fza.h>
> >> +#include <string-fzb.h>
> >> +#include <string-fzi.h>
> >> +#include <string-maskoff.h>
> >> +#include <string-opthr.h>
> >>  #include <string.h>
> >>
> >> -#include <stddef.h>
> >> +#undef memchr
> >>
> >> -#include <limits.h>
> >> -
> >> -#undef __memchr
> >> -#ifdef _LIBC
> >> -# undef memchr
> >> +#ifdef MEMCHR
> >> +# define __memchr MEMCHR
> >>  #endif
> >>
> >> -#ifndef weak_alias
> >> -# define __memchr memchr
> >> -#endif
> >> -
> >> -#ifndef MEMCHR
> >> -# define MEMCHR __memchr
> >> -#endif
> >> +static inline const char *
> >> +sadd (uintptr_t x, uintptr_t y)
> >> +{
> >> +  uintptr_t ret = INT_ADD_OVERFLOW (x, y) ? (uintptr_t)-1 : x + y;
> >> +  return (const char *)ret;
> >> +}
> >>
> >>  /* Search no more than N bytes of S for C.  */
> >>  void *
> >> -MEMCHR (void const *s, int c_in, size_t n)
> >> +__memchr (void const *s, int c_in, size_t n)
> >>  {
> >> -  /* On 32-bit hardware, choosing longword to be a 32-bit unsigned
> >> -     long instead of a 64-bit uintmax_t tends to give better
> >> -     performance.  On 64-bit hardware, unsigned long is generally 64
> >> -     bits already.  Change this typedef to experiment with
> >> -     performance.  */
> >> -  typedef unsigned long int longword;
> >> +  if (__glibc_unlikely (n == 0))
> >> +    return NULL;
> >>
> >> -  const unsigned char *char_ptr;
> >> -  const longword *longword_ptr;
> >> -  longword repeated_one;
> >> -  longword repeated_c;
> >> -  unsigned char c;
> >> +  uintptr_t s_int = (uintptr_t) s;
> >>
> >> -  c = (unsigned char) c_in;
> >> +  /* Set up a word, each of whose bytes is C.  */
> >> +  op_t repeated_c = repeat_bytes (c_in);
> >> +  op_t before_mask = create_mask (s_int);
> >>
> >> -  /* Handle the first few bytes by reading one byte at a time.
> >> -     Do this until CHAR_PTR is aligned on a longword boundary.  */
> >> -  for (char_ptr = (const unsigned char *) s;
> >> -       n > 0 && (size_t) char_ptr % sizeof (longword) != 0;
> >> -       --n, ++char_ptr)
> >> -    if (*char_ptr == c)
> >> -      return (void *) char_ptr;
> >> +  /* Compute the address of the last byte taking in consideration possible
> >> +     overflow.  */
> >> +  const char *lbyte = sadd (s_int, n - 1);
> >>
> >> -  longword_ptr = (const longword *) char_ptr;
> >> +  /* Compute the address of the word containing the last byte. */
> >> +  const op_t *lword = word_containing (lbyte);
> >>
> >> -  /* All these elucidatory comments refer to 4-byte longwords,
> >> -     but the theory applies equally well to any size longwords.  */
> >> +  /* Read the first word, but munge it so that bytes before the array
> >> +     will not match goal.  */
> >> +  const op_t *word_ptr = word_containing (s);
> >> +  op_t word = (*word_ptr | before_mask) ^ (repeated_c & before_mask);
> >>
> >> -  /* Compute auxiliary longword values:
> >> -     repeated_one is a value which has a 1 in every byte.
> >> -     repeated_c has c in every byte.  */
> >> -  repeated_one = 0x01010101;
> >> -  repeated_c = c | (c << 8);
> >> -  repeated_c |= repeated_c << 16;
> >> -  if (0xffffffffU < (longword) -1)
> >> +  while (has_eq (word, repeated_c) == 0)
> >>      {
> >> -      repeated_one |= repeated_one << 31 << 1;
> >> -      repeated_c |= repeated_c << 31 << 1;
> >> -      if (8 < sizeof (longword))
> >> -       {
> >> -         size_t i;
> >> -
> >> -         for (i = 64; i < sizeof (longword) * 8; i *= 2)
> >> -           {
> >> -             repeated_one |= repeated_one << i;
> >> -             repeated_c |= repeated_c << i;
> >> -           }
> >> -       }
> >> +      if (word_ptr == lword)
> >> +       return NULL;
> > Inuitively making lword, lword - 1 so that normal returns don't need the extra
> > null check would be faster.
>
> Hum, I did not follow; could you explain it with more details what you mean here?

I was thinking something like:

```
  op_t word = *word_ptr;
  op_t mask = find_eq_low (word, repeated_c)
      >> (CHAR_BIT * (s_int % sizeof (uintptr_t)));
  if (mask)
    {
      char *ret = (char *) s + index_first_ (mask);
      return (ret <= lbyte) ? ret : NULL;
    }
  if (word_ptr == lword)
    return NULL;

  word = *++word_ptr;
  while (word_ptr != lword)
    {
      if (has_eq (word, repeated_c))
return (char *) word_ptr + index_first_eq (word, repeated_c);
      word = *++word_ptr;
    }

  if (has_eq (word, repeated_c))
    {

      /* We found a match, but it might be in a byte past the end
of the array.  */
      char *ret = (char *) word_ptr + index_first_eq (word, repeated_c);
      if (ret <= lbyte)
return ret;
    }
  return NULL;
```

The idea is until the last byte you don't need the extra bounds check (tested
on test-memchr.c on little-endian).