Simplify and speedup strstr/strcasestr first match

Adhemerval Zanella wrote:

> +
> +  /* Determine length of NEEDLE and handle empty NEEDLE special case.  */
> +  if (needle[0] == '\0')
> +    return (char *) haystack;

> Maybe worth a __glibc_unlikely here?

I don't see any differences with __glibc_unlikely. GCC already does the right thing,
so I'd rather not clutter the code without a benefit.

> +  needle_len = strlen (needle);
> +
> +  /* Ensure HAYSTACK length is at least as long as NEEDLE length.  */
> +  haystack_len = __strnlen (haystack, needle_len + 256);
> +  if (haystack_len < needle_len)

> I would add a comment why 256 was picked, maybe adding to either avoid
> excessive strnlen calls or add extra memory access to check end of
> string on AVAILABLE.

Fair point - I improved the comment and the description. The chosen value gives a
good speedup on bench-strstr, so the readahead is useful but still small enough it
won't cause performance degradation in the case where there is an early match.
Here is the improved version:



Looking at the benchtests, both strstr and strcasestr spend a lot of time
in a slow initialization loop handling one character per iteration.
This can be simplified and use the much faster strlen/strnlen/strchr/memcmp.
Read ahead a few cachelines to reduce the number of strnlen calls, which
improves performance by ~3-4%.  This patch improves the time taken for the
full strstr benchtest by >40%.

Passes GLIBC regression tests.

ChangeLog:
2018-08-01  Wilco Dijkstra  <wdijkstr@arm.com>

        * string/strcasestr.c (STRCASESTR): Simplify and speedup first match.
        * string/strstr.c (AVAILABLE): Likewise.
--

On 01/08/2018 12:58, Wilco Dijkstra wrote:
> Adhemerval Zanella wrote:
> 
>> +
>> +  /* Determine length of NEEDLE and handle empty NEEDLE special case.  */
>> +  if (needle[0] == '\0')
>> +    return (char *) haystack;
> 
>> Maybe worth a __glibc_unlikely here?
> 
> I don't see any differences with __glibc_unlikely. GCC already does the right thing,
> so I'd rather not clutter the code without a benefit.
> 
>> +  needle_len = strlen (needle);
>> +
>> +  /* Ensure HAYSTACK length is at least as long as NEEDLE length.  */
>> +  haystack_len = __strnlen (haystack, needle_len + 256);
>> +  if (haystack_len < needle_len)
> 
>> I would add a comment why 256 was picked, maybe adding to either avoid
>> excessive strnlen calls or add extra memory access to check end of
>> string on AVAILABLE.
> 
> Fair point - I improved the comment and the description. The chosen value gives a
> good speedup on bench-strstr, so the readahead is useful but still small enough it
> won't cause performance degradation in the case where there is an early match.
> Here is the improved version:
> 
> 
> 
> Looking at the benchtests, both strstr and strcasestr spend a lot of time
> in a slow initialization loop handling one character per iteration.
> This can be simplified and use the much faster strlen/strnlen/strchr/memcmp.
> Read ahead a few cachelines to reduce the number of strnlen calls, which
> improves performance by ~3-4%.  This patch improves the time taken for the
> full strstr benchtest by >40%.
> 
> Passes GLIBC regression tests.
> 
> ChangeLog:
> 2018-08-01  Wilco Dijkstra  <wdijkstr@arm.com>
> 
>         * string/strcasestr.c (STRCASESTR): Simplify and speedup first match.
>         * string/strstr.c (AVAILABLE): Likewise.

LGTM, thanks.

> --
> 
> diff --git a/string/strcasestr.c b/string/strcasestr.c
> index 5909fe3cdba88e476c7a989a020f3611bbfeb1de..1f6b7b846f81d0d8fe77b390062d2d86be11b056 100644
> --- a/string/strcasestr.c
> +++ b/string/strcasestr.c
> @@ -58,31 +58,22 @@
>     case-insensitive comparison.  This function gives unspecified
>     results in multibyte locales.  */
>  char *
> -STRCASESTR (const char *haystack_start, const char *needle_start)
> +STRCASESTR (const char *haystack, const char *needle)
>  {
> -  const char *haystack = haystack_start;
> -  const char *needle = needle_start;
>    size_t needle_len; /* Length of NEEDLE.  */
>    size_t haystack_len; /* Known minimum length of HAYSTACK.  */
> -  bool ok = true; /* True if NEEDLE is prefix of HAYSTACK.  */
> -
> -  /* Determine length of NEEDLE, and in the process, make sure
> -     HAYSTACK is at least as long (no point processing all of a long
> -     NEEDLE if HAYSTACK is too short).  */
> -  while (*haystack && *needle)
> -    {
> -      ok &= (TOLOWER ((unsigned char) *haystack)
> -            == TOLOWER ((unsigned char) *needle));
> -      haystack++;
> -      needle++;
> -    }
> -  if (*needle)
> +
> +  /* Handle empty NEEDLE special case.  */
> +  if (needle[0] == '\0')
> +    return (char *) haystack;
> +
> +  /* Ensure HAYSTACK length is at least as long as NEEDLE length.
> +     Since a match may occur early on in a huge HAYSTACK, use strnlen
> +     and read ahead a few cachelines for improved performance.  */
> +  needle_len = strlen (needle);
> +  haystack_len = __strnlen (haystack, needle_len + 256);
> +  if (haystack_len < needle_len)
>      return NULL;
> -  if (ok)
> -    return (char *) haystack_start;
> -  needle_len = needle - needle_start;
> -  haystack = haystack_start + 1;
> -  haystack_len = needle_len - 1;
>  
>    /* Perform the search.  Abstract memory is considered to be an array
>       of 'unsigned char' values, not an array of 'char' values.  See
> @@ -90,10 +81,10 @@ STRCASESTR (const char *haystack_start, const char *needle_start)
>    if (needle_len < LONG_NEEDLE_THRESHOLD)
>      return two_way_short_needle ((const unsigned char *) haystack,
>                                   haystack_len,
> -                                (const unsigned char *) needle_start,
> +                                (const unsigned char *) needle,
>                                   needle_len);
>    return two_way_long_needle ((const unsigned char *) haystack, haystack_len,
> -                             (const unsigned char *) needle_start,
> +                             (const unsigned char *) needle,
>                                needle_len);
>  }
>  
> diff --git a/string/strstr.c b/string/strstr.c
> index 265e9f310ce507ce63740cc42d8ceea1d28dab01..33acdc5442d3e9031298a56e4f52a19e2ffa7d84 100644
> --- a/string/strstr.c
> +++ b/string/strstr.c
> @@ -50,33 +50,32 @@
>     if NEEDLE is empty, otherwise NULL if NEEDLE is not found in
>     HAYSTACK.  */
>  char *
> -STRSTR (const char *haystack_start, const char *needle_start)
> +STRSTR (const char *haystack, const char *needle)
>  {
> -  const char *haystack = haystack_start;
> -  const char *needle = needle_start;
>    size_t needle_len; /* Length of NEEDLE.  */
>    size_t haystack_len; /* Known minimum length of HAYSTACK.  */
> -  bool ok = true; /* True if NEEDLE is prefix of HAYSTACK.  */
> -
> -  /* Determine length of NEEDLE, and in the process, make sure
> -     HAYSTACK is at least as long (no point processing all of a long
> -     NEEDLE if HAYSTACK is too short).  */
> -  while (*haystack && *needle)
> -    ok &= *haystack++ == *needle++;
> -  if (*needle)
> +
> +  /* Handle empty NEEDLE special case.  */
> +  if (needle[0] == '\0')
> +    return (char *) haystack;
> +
> +  /* Skip until we find the first matching char from NEEDLE.  */
> +  haystack = strchr (haystack, needle[0]);
> +  if (haystack == NULL || needle[1] == '\0')
> +    return (char *) haystack;
> +
> +  /* Ensure HAYSTACK length is at least as long as NEEDLE length.
> +     Since a match may occur early on in a huge HAYSTACK, use strnlen
> +     and read ahead a few cachelines for improved performance.  */
> +  needle_len = strlen (needle);
> +  haystack_len = __strnlen (haystack, needle_len + 256);
> +  if (haystack_len < needle_len)
>      return NULL;
> -  if (ok)
> -    return (char *) haystack_start;
> -
> -  /* Reduce the size of haystack using strchr, since it has a smaller
> -     linear coefficient than the Two-Way algorithm.  */
> -  needle_len = needle - needle_start;
> -  haystack = strchr (haystack_start + 1, *needle_start);
> -  if (!haystack || __builtin_expect (needle_len == 1, 0))
> +
> +  /* Check whether we have a match.  This improves performance since we avoid
> +     the initialization overhead of the two-way algorithm.  */
> +  if (memcmp (haystack, needle, needle_len) == 0)
>      return (char *) haystack;
> -  needle -= needle_len;
> -  haystack_len = (haystack > haystack_start + needle_len ? 1
> -                 : needle_len + haystack_start - haystack);
>  
>    /* Perform the search.  Abstract memory is considered to be an array
>       of 'unsigned char' values, not an array of 'char' values.  See
> 
> 
>         
>