tilegx: provide optimized strnlen, strstr, and strcasestr
diff mbox

Message ID 201410021559.s92Fx8CN020856@farm-0002.internal.tilera.com
State New
Headers show

Commit Message

Chris Metcalf Sept. 16, 2014, 12:10 a.m. UTC
strnlen() is based on the existing tile strlen() with length
checking added.  It speeds up by up to 5x, but on average across
the benchtest corpus by around 35%.  No regressions are seen.

strstr() does 8-byte aligned loads and compares using a 2-byte
filter on the first two bytes of the needle and then testing
the remaining bytes in needle using memcmp().  It speeds up
about 5x in the best case (for "found" needles), about 2x looking
at benchtest as a whole, with some slowdowns as much as 45%.
on a few cases (including the "fail" case for 128KB search).

strcasestr() is based on strstr() but uses a SIMD tolower
routine to convert 8-bytes to lower case in 5 instructions.
It also uses a 2-byte filter and then strncasecmp() for the
remaining bytes.  strncasecmp() is not optimized for SIMD, so
there is futher room for improvement.  However, it is still up
to 16x faster for "found" needles, averaging 2x faster on the
whole corpus of benchtests.  It does slow down by up to 35%
on a few cases, similarly to strstr().
---
sysdeps/tile/tilegx/strcasestr.c    |  55 ++++++++
sysdeps/tile/tilegx/string-endian.h |  22 ++-
sysdeps/tile/tilegx/strnlen.c       |  58 ++++++++
sysdeps/tile/tilegx/strstr.c        | 271 ++++++++++++++++++++++++++++++++++++
4 files changed, 401 insertions(+), 5 deletions(-)
create mode 100644 sysdeps/tile/tilegx/strcasestr.c
create mode 100644 sysdeps/tile/tilegx/strnlen.c
create mode 100644 sysdeps/tile/tilegx/strstr.c

2014-10-02  Chris Metcalf  <cmetcalf@tilera.com>

	* sysdeps/tile/tilegx/string-endian.h (STRSHIFT): New macro.
	* sysdeps/tile/tilegx/strcasestr.c: New file.
	* sysdeps/tile/tilegx/strnlen.c: New file.
	* sysdeps/tile/tilegx/strstr.c: New file.

Comments

Joseph Myers Oct. 2, 2014, 6:09 p.m. UTC | #1
On Mon, 15 Sep 2014, Chris Metcalf wrote:

> +   Copyright (C) 1994, 1996-2000, 2004, 2008, 2009
> +   Free Software Foundation, Inc.

All FSF copyright notices should now use ranges ending with the current 
year (either <year>-2014, or just 2014), unless there is a pre-1991 gap in 
the copyright years (a few files special-cased in 
scripts/update-copyrights).

> +   You should have received a copy of the GNU Lesser General Public
> +   License along with the GNU C Library; if not, write to the Free
> +   Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
> +   02111-1307 USA.  */

They should also use URLs, not very out of date postal addresses.

> +/* Copyright (C) 2013 Free Software Foundation, Inc.
> +   This file is part of the GNU C Library.
> +   Contributed by Chris Metcalf <cmetcalf@tilera.com>, 2011.

New files should also not have "Contributed by" lines.
Chris Metcalf Oct. 2, 2014, 7:13 p.m. UTC | #2
Grrr - I fixed all of these the first time I imported these changes, but then later in the week I re-imported these files from our (2.12-based) shipping code repository and forgot to re-do it.  I will fix them now - I don't think it merits another reposting, though I'm happy to if anyone would like.

Thanks.

On 10/2/2014 2:09 PM, Joseph S. Myers wrote:
> On Mon, 15 Sep 2014, Chris Metcalf wrote:
>
>> +   Copyright (C) 1994, 1996-2000, 2004, 2008, 2009
>> +   Free Software Foundation, Inc.
> All FSF copyright notices should now use ranges ending with the current
> year (either <year>-2014, or just 2014), unless there is a pre-1991 gap in
> the copyright years (a few files special-cased in
> scripts/update-copyrights).
>
>> +   You should have received a copy of the GNU Lesser General Public
>> +   License along with the GNU C Library; if not, write to the Free
>> +   Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
>> +   02111-1307 USA.  */
> They should also use URLs, not very out of date postal addresses.
>
>> +/* Copyright (C) 2013 Free Software Foundation, Inc.
>> +   This file is part of the GNU C Library.
>> +   Contributed by Chris Metcalf <cmetcalf@tilera.com>, 2011.
> New files should also not have "Contributed by" lines.
>
Joseph Myers Oct. 6, 2014, 5:33 p.m. UTC | #3
On Thu, 2 Oct 2014, Chris Metcalf wrote:

> Grrr - I fixed all of these the first time I imported these changes, but then
> later in the week I re-imported these files from our (2.12-based) shipping
> code repository and forgot to re-do it.  I will fix them now - I don't think
> it merits another reposting, though I'm happy to if anyone would like.

It appears you've pushed additions of several files with these problems 
still present.
Ondřej Bílka June 3, 2015, 8:26 a.m. UTC | #4
On Mon, Sep 15, 2014 at 08:10:18PM -0400, Chris Metcalf wrote:
> strnlen() is based on the existing tile strlen() with length
> checking added.  It speeds up by up to 5x, but on average across
> the benchtest corpus by around 35%.  No regressions are seen.
> 
> strstr() does 8-byte aligned loads and compares using a 2-byte
> filter on the first two bytes of the needle and then testing
> the remaining bytes in needle using memcmp().  It speeds up
> about 5x in the best case (for "found" needles), about 2x looking
> at benchtest as a whole, with some slowdowns as much as 45%.
> on a few cases (including the "fail" case for 128KB search).
> 
> strcasestr() is based on strstr() but uses a SIMD tolower
> routine to convert 8-bytes to lower case in 5 instructions.
> It also uses a 2-byte filter and then strncasecmp() for the
> remaining bytes.  strncasecmp() is not optimized for SIMD, so
> there is futher room for improvement.  However, it is still up
> to 16x faster for "found" needles, averaging 2x faster on the
> whole corpus of benchtests.  It does slow down by up to 35%
> on a few cases, similarly to strstr().
> ---
> sysdeps/tile/tilegx/strcasestr.c    |  55 ++++++++
> sysdeps/tile/tilegx/string-endian.h |  22 ++-
> sysdeps/tile/tilegx/strnlen.c       |  58 ++++++++
> sysdeps/tile/tilegx/strstr.c        | 271 ++++++++++++++++++++++++++++++++++++
> 4 files changed, 401 insertions(+), 5 deletions(-)
> create mode 100644 sysdeps/tile/tilegx/strcasestr.c
> create mode 100644 sysdeps/tile/tilegx/strnlen.c
> create mode 100644 sysdeps/tile/tilegx/strstr.c
> 
I didn't notice this thread before so didn't comment.

First there is bug in strcasestr, as you couldn't always use vector ascii
conversion, you would need to check that with:

  __locale_t loc = _NL_CURRENT_LOCALE;
  struct __locale_data *ctype = loc->__locales[LC_CTYPE];
  int nonascii = ctype->values[_NL_ITEM_INDEX(_NL_CTYPE_NONASCII_CASE)].word;

But you don't need vector conversion there. Just do comparisons with
tolower(x) and toupper(x). I just realized that my strcasestr was
overcomplicated as I assumed that many characters could have same
tolower(x).

Best course of action would be wait until I merge my strstr skeleton and
you map instrincs. Then we could delete these. You use same idea,
skeleton adds many technical speedups like that there was bottleneck in
checking last character alone instead of as digraph which is solved by
merging last two loads to get full word with characters.
Chris Metcalf June 9, 2015, 7:45 p.m. UTC | #5
On 06/03/2015 04:26 AM, Ondřej Bílka wrote:
> On Mon, Sep 15, 2014 at 08:10:18PM -0400, Chris Metcalf wrote:
>> strnlen() is based on the existing tile strlen() with length
>> checking added.  It speeds up by up to 5x, but on average across
>> the benchtest corpus by around 35%.  No regressions are seen.
>>
>> strstr() does 8-byte aligned loads and compares using a 2-byte
>> filter on the first two bytes of the needle and then testing
>> the remaining bytes in needle using memcmp().  It speeds up
>> about 5x in the best case (for "found" needles), about 2x looking
>> at benchtest as a whole, with some slowdowns as much as 45%.
>> on a few cases (including the "fail" case for 128KB search).
>>
>> strcasestr() is based on strstr() but uses a SIMD tolower
>> routine to convert 8-bytes to lower case in 5 instructions.
>> It also uses a 2-byte filter and then strncasecmp() for the
>> remaining bytes.  strncasecmp() is not optimized for SIMD, so
>> there is futher room for improvement.  However, it is still up
>> to 16x faster for "found" needles, averaging 2x faster on the
>> whole corpus of benchtests.  It does slow down by up to 35%
>> on a few cases, similarly to strstr().
>> ---
>> sysdeps/tile/tilegx/strcasestr.c    |  55 ++++++++
>> sysdeps/tile/tilegx/string-endian.h |  22 ++-
>> sysdeps/tile/tilegx/strnlen.c       |  58 ++++++++
>> sysdeps/tile/tilegx/strstr.c        | 271 ++++++++++++++++++++++++++++++++++++
>> 4 files changed, 401 insertions(+), 5 deletions(-)
>> create mode 100644 sysdeps/tile/tilegx/strcasestr.c
>> create mode 100644 sysdeps/tile/tilegx/strnlen.c
>> create mode 100644 sysdeps/tile/tilegx/strstr.c
>>
> I didn't notice this thread before so didn't comment.
>
> First there is bug in strcasestr, as you couldn't always use vector ascii
> conversion, you would need to check that with:
>
>    __locale_t loc = _NL_CURRENT_LOCALE;
>    struct __locale_data *ctype = loc->__locales[LC_CTYPE];
>    int nonascii = ctype->values[_NL_ITEM_INDEX(_NL_CTYPE_NONASCII_CASE)].word;
>
> But you don't need vector conversion there. Just do comparisons with
> tolower(x) and toupper(x). I just realized that my strcasestr was
> overcomplicated as I assumed that many characters could have same
> tolower(x).
>
> Best course of action would be wait until I merge my strstr skeleton and
> you map instrincs. Then we could delete these. You use same idea,
> skeleton adds many technical speedups like that there was bottleneck in
> checking last character alone instead of as digraph which is solved by
> merging last two loads to get full word with characters.

Thanks for spotting the bug.  I've filed it as bug 18510.  I'll
hold off trying to do a point fix until it becomes clear whether
or not your more general fixes will hit mainline.

Patch
diff mbox

diff --git a/sysdeps/tile/tilegx/strcasestr.c b/sysdeps/tile/tilegx/strcasestr.c
new file mode 100644
index 000000000000..13b0a8485890
--- /dev/null
+++ b/sysdeps/tile/tilegx/strcasestr.c
@@ -0,0 +1,55 @@ 
+/* Return the offset of one string within another.
+   Copyright (C) 1994, 1996-2000, 2004, 2008, 2009
+   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, write to the Free
+   Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+   02111-1307 USA.  */
+
+#if HAVE_CONFIG_H
+# include <config.h>
+#endif
+
+/* Specification.  */
+#include <string.h>
+
+#include <ctype.h>
+#include <stdbool.h>
+#include <strings.h>
+
+#define USE_AS_STRCASESTR
+#define STRSTR __strcasestr
+#define STRSTR2 strcasestr2
+#define STRCHR strcasechr
+#define STRSTR_SCAN strcasestr_scan
+
+#undef strcasestr
+#undef __strcasestr
+
+#ifndef STRCASESTR
+#define STRCASESTR __strcasestr
+#endif
+
+#define TOLOWER(Ch) (isupper (Ch) ? tolower (Ch) : (Ch))
+
+#define CANON_ELEMENT(c) TOLOWER (c)
+#define CMP_FUNC(p1, p2, l)				\
+  __strncasecmp ((const char *) (p1), (const char *) (p2), l)
+
+#include "strstr.c"
+
+#ifndef NO_ALIAS
+weak_alias (__strcasestr, strcasestr)
+#endif
diff --git a/sysdeps/tile/tilegx/string-endian.h b/sysdeps/tile/tilegx/string-endian.h
index 47333891e072..2dbc1e4a4fe1 100644
--- a/sysdeps/tile/tilegx/string-endian.h
+++ b/sysdeps/tile/tilegx/string-endian.h
@@ -16,24 +16,36 @@ 
    License along with the GNU C Library.  If not, see
    <http://www.gnu.org/licenses/>.  */
 
-/* Provide a mask based on the pointer alignment that
+#include <endian.h>
+#include <stdint.h>
+
+/* Provide a set of macros to help keep endianness #ifdefs out of
+   the string functions.
+
+   MASK: Provide a mask based on the pointer alignment that
    sets up non-zero bytes before the beginning of the string.
    The MASK expression works because shift counts are taken mod 64.
-   Also, specify how to count "first" and "last" bits
-   when the bits have been read as a word.  */
 
-#include <stdint.h>
+   NULMASK: Clear bytes beyond a given point in the string.
+
+   CFZ: Find the first zero bit in the 8 string bytes in a long.
+
+   REVCZ: Find the last zero bit in the 8 string bytes in a long.
+
+   STRSHIFT: Shift N bits towards the start of the string.  */
 
-#ifndef __BIG_ENDIAN__
+#if __BYTE_ORDER == __LITTLE_ENDIAN
 #define MASK(x) (__insn_shl(1ULL, (x << 3)) - 1)
 #define NULMASK(x) ((2ULL << x) - 1)
 #define CFZ(x) __insn_ctz(x)
 #define REVCZ(x) __insn_clz(x)
+#define STRSHIFT(x,n) ((x) >> n)
 #else
 #define MASK(x) (__insn_shl(-2LL, ((-x << 3) - 1)))
 #define NULMASK(x) (-2LL << (63 - x))
 #define CFZ(x) __insn_clz(x)
 #define REVCZ(x) __insn_ctz(x)
+#define STRSHIFT(x,n) ((x) << n)
 #endif
 
 /* Create eight copies of the byte in a uint64_t.  Byte Shuffle uses
diff --git a/sysdeps/tile/tilegx/strnlen.c b/sysdeps/tile/tilegx/strnlen.c
new file mode 100644
index 000000000000..33ecc033f6ac
--- /dev/null
+++ b/sysdeps/tile/tilegx/strnlen.c
@@ -0,0 +1,58 @@ 
+/* Copyright (C) 2013 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+   Contributed by Chris Metcalf <cmetcalf@tilera.com>, 2011.
+
+   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, write to the Free
+   Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+   02111-1307 USA.  */
+
+#include <string.h>
+#include <stdint.h>
+#include "string-endian.h"
+
+/* Find the length of S, but scan at most MAXLEN characters.  If no
+   '\0' terminator is found in that many characters, return MAXLEN.  */
+size_t
+__strnlen (const char *s, size_t maxlen)
+{
+  /* When maxlen is 0, can't read any bytes or it might cause a page fault.  */
+  if (maxlen == 0)
+    return 0;
+
+  /* Get an aligned pointer. */
+  const uintptr_t s_int = (uintptr_t) s;
+  const uint64_t *p = (const uint64_t *) (s_int & -8);
+  size_t bytes_read = sizeof (*p) - (s_int & (sizeof (*p) - 1));
+
+  /* Read and MASK the first word. */
+  uint64_t v = *p | MASK (s_int);
+
+  uint64_t bits;
+  while ((bits = __insn_v1cmpeqi (v, 0)) == 0)
+    {
+      if (bytes_read >= maxlen)
+	{
+	  /* Read maxlen bytes and didn't find the terminator. */
+	  return maxlen;
+	}
+      v = *++p;
+      bytes_read += sizeof (v);
+    }
+
+  /* Found '\0', check it is not larger than maxlen */
+  size_t len = ((const char *) p) + (CFZ (bits) >> 3) - s;
+  return (len < maxlen ? len : maxlen);
+}
+weak_alias (__strnlen, strnlen)
+libc_hidden_def (strnlen)
diff --git a/sysdeps/tile/tilegx/strstr.c b/sysdeps/tile/tilegx/strstr.c
new file mode 100644
index 000000000000..2627e758c739
--- /dev/null
+++ b/sysdeps/tile/tilegx/strstr.c
@@ -0,0 +1,271 @@ 
+/* strstr with Tile intrinsics
+   Copyright (C) 2013 Free Software Foundation, Inc.
+   Contributed by Intel Corporation.
+   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, write to the Free
+   Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+   02111-1307 USA.  */
+
+/* Specification of strstr.  */
+#include <string.h>
+
+#include <stdbool.h>
+#include "string-endian.h"
+
+#define RETURN_TYPE char *
+#define AVAILABLE(h, h_l, j, n_l)			\
+  (!memchr ((h) + (h_l), '\0', (j) + (n_l) - (h_l))	\
+   && ((h_l) = (j) + (n_l)))
+#include "str-two-way.h"
+
+#undef strstr
+
+#ifndef STRSTR
+#define STRSTR strstr
+#endif
+
+#ifndef STRSTR2
+#define STRSTR2 strstr2
+#endif
+
+#ifndef STRCHR
+#define STRCHR strchr
+#endif
+
+#ifndef STRSTR_SCAN
+#define STRSTR_SCAN strstr_scan
+#endif
+
+#ifndef TOLOWER
+# define TOLOWER(Ch) (Ch)
+#endif
+
+#ifdef USE_AS_STRCASESTR
+
+static uint64_t
+vec_tolower (uint64_t cc)
+{
+  /* For Uppercases letters, add 32 to convert to lower case.  */
+  uint64_t less_than_eq_Z = __insn_v1cmpltui (cc, 'Z' + 1);
+  uint64_t less_than_A =  __insn_v1cmpltui (cc, 'A');
+  uint64_t is_upper = __insn_v1cmpne (less_than_eq_Z, less_than_A);
+  return __insn_v1add (cc,__insn_v1shli (is_upper, 5));
+}
+
+/* There is no strcasechr() defined, but needed for 1 byte case
+   of strcasestr(), so create it here.  */
+
+static char *
+strcasechr (const char *s, int c)
+{
+  int z, g;
+
+  c = tolower (c);
+
+  /* Get an aligned pointer.  */
+  const uintptr_t s_int = (uintptr_t) s;
+  const uint64_t *p = (const uint64_t *) (s_int & -8);
+
+  /* Create eight copies of the byte for which we are looking.  */
+  const uint64_t goal = copy_byte(c);
+
+  /* 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).  */
+  const uint64_t before_mask = MASK (s_int);
+  uint64_t v =
+    (vec_tolower (*p) | before_mask) ^ (goal & __insn_v1shrui (before_mask, 1));
+
+  uint64_t zero_matches, goal_matches;
+  while (1)
+    {
+      /* Look for a terminating '\0'.  */
+      zero_matches = __insn_v1cmpeqi (v, 0);
+
+      /* Look for the goal byte.  */
+      goal_matches = __insn_v1cmpeq (v, goal);
+
+      if (__builtin_expect ((zero_matches | goal_matches) != 0, 0))
+        break;
+
+      v = vec_tolower (*++p);
+    }
+
+  z = CFZ (zero_matches);
+  g = CFZ (goal_matches);
+
+  /* If we found c before '\0' we got a match. Note that if c == '\0'
+     then g == z, and we correctly return the address of the '\0'
+     rather than NULL.  */
+  return (g <= z) ? ((char *) p) + (g >> 3) : NULL;
+}
+
+# define vec_load(p) vec_tolower (*(p))
+# define STRCHR strcasechr
+# define CMP_FUNC __strncasecmp
+
+#else
+
+# define vec_load(p) (*(p))
+# define STRCHR strchr
+# define CMP_FUNC memcmp
+
+#endif
+
+
+/* Compare 2-character needle using SIMD.  */
+static char *
+STRSTR2 (const char *haystack_start, const char *needle)
+{
+  int z, g;
+
+  __insn_prefetch (haystack_start + 64);
+
+  /* Get an aligned pointer.  */
+  const uintptr_t s_int = (uintptr_t) haystack_start;
+  const uint64_t *p = (const uint64_t *) (s_int & -8);
+
+  /* Create eight copies of the first byte for which we are looking.  */
+  const uint64_t byte1 = copy_byte (TOLOWER (*needle));
+  /* Create eight copies of the second byte for which we are looking.  */
+  const uint64_t byte2 = copy_byte (TOLOWER (*(needle + 1)));
+
+  /* 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).  */
+  const uint64_t before_mask = MASK (s_int);
+  uint64_t v =
+    (vec_load (p) | before_mask) ^ (byte1 & __insn_v1shrui (before_mask, 1));
+
+  uint64_t zero_matches, goal_matches;
+  while (1)
+    {
+      /* Look for a terminating '\0'.  */
+      zero_matches = __insn_v1cmpeqi (v, 0);
+      uint64_t byte1_matches = __insn_v1cmpeq (v, byte1);
+      if (__builtin_expect (zero_matches, 0))
+	{
+	  /* This is the last vector.  Don't worry about matches
+	     crossing into the next vector.  Shift the second byte
+	     back 1 byte to align it with the first byte, then and to
+	     check for both matching.  Each vector has a 1 in the LSB
+	     of the byte if there was match.  */
+	  uint64_t byte2_matches = __insn_v1cmpeq (v, byte2);
+	  goal_matches = byte1_matches & STRSHIFT (byte2_matches, 8);
+	  break;
+	}
+      else
+	{
+	  /* This is not the last vector, so load the next vector now.
+	     And compare byte2 to the 8-bytes starting 1 byte shifted from v,
+	     which goes 1-byte into the next vector.  */
+	  uint64_t v2 = vec_load (p + 1);
+	  if (byte1_matches)
+	    {
+	      /* 8-bytes starting 1 byte into v.  */
+	      v = __insn_dblalign (v, v2, (void*)1);
+	      uint64_t byte2_matches_shifted = __insn_v1cmpeq (v, byte2);
+	      goal_matches = byte1_matches & byte2_matches_shifted;
+	      if (__builtin_expect (goal_matches != 0, 0))
+		break;
+	    }
+	  __insn_prefetch (p + 4);
+	  /* Move to next vector.  */
+	  v = v2;
+	  p++;
+	}
+    }
+
+  z = CFZ (zero_matches);
+  g = CFZ (goal_matches);
+
+  /* If we found the match before '\0' we got a true match. Note that
+     if c == '\0' then g == z, and we correctly return the address of
+     the '\0' rather than NULL.  */
+  return (g <= z) ? ((char *) p) + (g >> 3) : NULL;
+}
+
+/* Scan for NEEDLE, using the first two characters as a filter.  */
+static char *
+STRSTR_SCAN (const char *haystack, const char *needle,
+	     unsigned int needle_len)
+{
+  char *match;
+  while (1)
+    {
+      match = STRSTR2 (haystack, needle);
+      if (match == NULL)
+	return NULL;
+      /* Found first two characters of needle, check for remainder.  */
+      if (CMP_FUNC (match + 2, needle + 2, needle_len - 2) == 0)
+	return match;
+      /* Move past the previous match. Could be +2 instead of +1 if
+         first two characters are different, but that tested slower.  */
+      haystack = match + 1;
+    }
+}
+
+/* Return the first occurrence of NEEDLE in HAYSTACK.  Return HAYSTACK
+   if NEEDLE is empty, otherwise NULL if NEEDLE is not found in
+   HAYSTACK.  */
+char *
+STRSTR (const char *haystack_start, const char *needle_start)
+{
+  const char *haystack = haystack_start;
+  const char *needle = needle_start;
+  __insn_prefetch (haystack);
+  size_t needle_len = strlen (needle_start); /* Length of NEEDLE.  */
+  size_t haystack_len; /* Known minimum length of HAYSTACK.  */
+
+  if (needle_len <= 2)
+    {
+      if (needle_len == 1)
+	return STRCHR (haystack_start, *needle_start);
+      if (needle_len == 0)
+	return (char *) haystack_start;
+      else
+	return STRSTR2 (haystack_start, needle_start);
+    }
+
+  /* Fail if NEEDLE is longer than HAYSTACK.  */
+  if (strnlen (haystack, needle_len) < needle_len)
+    return NULL;
+
+  /* Perform the search.  Abstract memory is considered to be an array
+     of 'unsigned char' values, not an array of 'char' values.  See
+     ISO C 99 section 6.2.6.1.  */
+  if (needle_len < 40)
+    return STRSTR_SCAN (haystack_start, needle_start, needle_len);
+  else
+    {
+      /* Reduce the size of haystack using STRSTR2, since it has a smaller
+	 linear coefficient than the Two-Way algorithm.  */
+      haystack = STRSTR2 (haystack_start, needle_start);
+      if (haystack == NULL)
+	return NULL;
+      needle = needle_start;
+      haystack_len = (haystack > haystack_start + needle_len ? 1
+		      : needle_len + haystack_start - haystack);
+
+      return two_way_long_needle ((const unsigned char *) haystack,
+				  haystack_len,
+				  (const unsigned char *) needle, needle_len);
+    }
+}
+#ifndef USE_AS_STRCASESTR
+libc_hidden_builtin_def (STRSTR)
+#endif
+
+#undef LONG_NEEDLE_THRESHOLD