Re: [PATCH v5 08/17] string: Improve generic strchrnul

[Adhemerval Zanella Netto <adhemerval.zanella@linaro.org>]

On 09/01/23 20:33, Richard Henderson wrote:
> On 1/9/23 12:35, Adhemerval Zanella Netto wrote:
>>
>>
>> On 05/01/23 20:17, Noah Goldstein wrote:
>>> On Mon, Sep 19, 2022 at 1:04 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} functions.
>>>>
>>>> Checked on x86_64-linux-gnu, i686-linux-gnu, powerpc64-linux-gnu,
>>>> and powerpc-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/strchrnul.c                            | 156 +++---------------
>>>>   .../power4/multiarch/strchrnul-ppc32.c        |   4 -
>>>>   sysdeps/s390/strchrnul-c.c                    |   2 -
>>>>   3 files changed, 24 insertions(+), 138 deletions(-)
>>>>
>>>> diff --git a/string/strchrnul.c b/string/strchrnul.c
>>>> index 0cc1fc6bb0..67defa3dab 100644
>>>> --- a/string/strchrnul.c
>>>> +++ b/string/strchrnul.c
>>>> @@ -1,10 +1,5 @@
>>>>   /* 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
>>>> -   bug fix and commentary by Jim Blandy (jimb@ai.mit.edu);
>>>> -   adaptation to strchr 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
>>>> @@ -21,146 +16,43 @@
>>>>      <https://www.gnu.org/licenses/>.  */
>>>>
>>>>   #include <string.h>
>>>> -#include <memcopy.h>
>>>>   #include <stdlib.h>
>>>> +#include <stdint.h>
>>>> +#include <string-fza.h>
>>>> +#include <string-fzb.h>
>>>> +#include <string-fzi.h>
>>>> +#include <string-maskoff.h>
>>>>
>>>>   #undef __strchrnul
>>>>   #undef strchrnul
>>>>
>>>> -#ifndef STRCHRNUL
>>>> -# define STRCHRNUL __strchrnul
>>>> +#ifdef STRCHRNUL
>>>> +# define __strchrnul STRCHRNUL
>>>>   #endif
>>>>
>>>>   /* Find the first occurrence of C in S or the final NUL byte.  */
>>>>   char *
>>>> -STRCHRNUL (const char *s, int c_in)
>>>> +__strchrnul (const char *str, int c_in)
>>>>   {
>>>> -  const unsigned char *char_ptr;
>>>> -  const unsigned long int *longword_ptr;
>>>> -  unsigned long int longword, magic_bits, charmask;
>>>> -  unsigned char c;
>>>> -
>>>> -  c = (unsigned char) c_in;
>>>> -
>>>> -  /* Handle the first few characters by reading one character at a time.
>>>> -     Do this until CHAR_PTR is aligned on a longword boundary.  */
>>>> -  for (char_ptr = (const unsigned char *) s;
>>>> -       ((unsigned long int) char_ptr & (sizeof (longword) - 1)) != 0;
>>>> -       ++char_ptr)
>>>> -    if (*char_ptr == c || *char_ptr == '\0')
>>>> -      return (void *) char_ptr;
>>>> -
>>>> -  /* All these elucidatory comments refer to 4-byte longwords,
>>>> -     but the theory applies equally well to 8-byte longwords.  */
>>>> -
>>>> -  longword_ptr = (unsigned long int *) char_ptr;
>>>> -
>>>> -  /* Bits 31, 24, 16, and 8 of this number are zero.  Call these bits
>>>> -     the "holes."  Note that there is a hole just to the left of
>>>> -     each byte, with an extra at the end:
>>>> -
>>>> -     bits:  01111110 11111110 11111110 11111111
>>>> -     bytes: AAAAAAAA BBBBBBBB CCCCCCCC DDDDDDDD
>>>> -
>>>> -     The 1-bits make sure that carries propagate to the next 0-bit.
>>>> -     The 0-bits provide holes for carries to fall into.  */
>>>> -  magic_bits = -1;
>>>> -  magic_bits = magic_bits / 0xff * 0xfe << 1 >> 1 | 1;
>>>> -
>>>> -  /* Set up a longword, each of whose bytes is C.  */
>>>> -  charmask = c | (c << 8);
>>>> -  charmask |= charmask << 16;
>>>> -  if (sizeof (longword) > 4)
>>>> -    /* Do the shift in two steps to avoid a warning if long has 32 bits.  */
>>>> -    charmask |= (charmask << 16) << 16;
>>>> -  if (sizeof (longword) > 8)
>>>> -    abort ();
>>>> -
>>>> -  /* Instead of the traditional loop which tests each character,
>>>> -     we will test a longword at a time.  The tricky part is testing
>>>> -     if *any of the four* bytes in the longword in question are zero.  */
>>>> -  for (;;)
>>>> -    {
>>>> -      /* We tentatively exit the loop if adding MAGIC_BITS to
>>>> -        LONGWORD fails to change any of the hole bits of LONGWORD.
>>>> -
>>>> -        1) Is this safe?  Will it catch all the zero bytes?
>>>> -        Suppose there is a byte with all zeros.  Any carry bits
>>>> -        propagating from its left will fall into the hole at its
>>>> -        least significant bit and stop.  Since there will be no
>>>> -        carry from its most significant bit, the LSB of the
>>>> -        byte to the left will be unchanged, and the zero will be
>>>> -        detected.
>>>> +  /* Set up a word, each of whose bytes is C.  */
>>>> +  op_t repeated_c = repeat_bytes (c_in);
>>>>
>>>> -        2) Is this worthwhile?  Will it ignore everything except
>>>> -        zero bytes?  Suppose every byte of LONGWORD has a bit set
>>>> -        somewhere.  There will be a carry into bit 8.  If bit 8
>>>> -        is set, this will carry into bit 16.  If bit 8 is clear,
>>>> -        one of bits 9-15 must be set, so there will be a carry
>>>> -        into bit 16.  Similarly, there will be a carry into bit
>>>> -        24.  If one of bits 24-30 is set, there will be a carry
>>>> -        into bit 31, so all of the hole bits will be changed.
>>>> +  /* Align the input address to op_t.  */
>>>> +  uintptr_t s_int = (uintptr_t) str;
>>>> +  const op_t *word_ptr = word_containing (str);
>>>>
>>>> -        The one misfire occurs when bits 24-30 are clear and bit
>>>> -        31 is set; in this case, the hole at bit 31 is not
>>>> -        changed.  If we had access to the processor carry flag,
>>>> -        we could close this loophole by putting the fourth hole
>>>> -        at bit 32!
>>>> +  /* Read the first aligned word, but force bytes before the string to
>>>> +     match neither zero nor goal (we make sure the high bit of each byte
>>>> +     is 1, and the low 7 bits are all the opposite of the goal byte).  */
>>>> +  op_t bmask = create_mask (s_int);
>>>> +  op_t word = (*word_ptr | bmask) ^ (repeated_c & highbit_mask (bmask));
>>>
>>> Think much clearer (and probably better codegen) is:
>>> find_zero_eq_low/all(word, repeated) >> (s_int * CHAR_BIT)
>>
>> It does not seem to work, at least not replacing the two lines with:
>>
>>    op_t word = find_zero_eq_all/low (*word_ptr, repeated_c) >> (s_int * CHAR_BIT);
> 
> Oh, two fine points:
> 
> (1) big-endian would want shifting left,
> (2) alpha would want shifting by bits not bytes,
>     because the cmpbge insn produces an 8-bit mask.
> 
> so you'd need to hide this shift in the headers like create_mask().

Alright, the following works:


static __always_inline op_t
check_mask (op_t word, uintptr_t s_int)
{
  if (__BYTE_ORDER == __LITTLE_ENDIAN)
    return word >> (CHAR_BIT * (s_int % sizeof (s_int)));
  else
    return word << (CHAR_BIT * (s_int % sizeof (s_int)));
}

char *
__strchrnul (const char *str, int c_in)
{
  op_t repeated_c = repeat_bytes (c_in);

  uintptr_t s_int = (uintptr_t) str;
  const op_t *word_ptr = word_containing (str);

  op_t word = *word_ptr;

  op_t mask = check_mask (find_zero_eq_all (word, repeated_c), s_int);
  if (mask != 0)
    return (char *) str + index_first_(mask);

  do
    word = *++word_ptr;
  while (! has_zero_eq (word, repeated_c));

  op_t found = index_first_zero_eq (word, repeated_c);
  return (char *) word_ptr + found;
}

 
I had to use find_zero_eq_all to avoid uninitialized bytes, that triggered
some regression on tests that use strchr (for instance test-strpbrk).

I will update the patch based on this version.