Re: memrchr speed

[Eric Blake]

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

According to Bruno Haible on 4/26/2008 4:38 AM:
| Eric Blake asked:
|> Also, is anyone interested in making gnulib's memchr and strchrnul more
|> efficient by copying the optimizations learned in memchr2?
|
| Checked in like this:
|
| 2008-04-26  Eric Blake  <address@hidden>
|             Bruno Haible  <address@hidden>
|
|       * lib/memchr.c: Include intprops.h.
|       (__memchr): Optimize parallel detection of matching bytes. Rename local
|       variables. Add explanatory comments.

And to memrchr like this:

- --
Don't work too hard, make some time for fun as well!

Eric Blake             address@hidden
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.9 (Cygwin)
Comment: Public key at home.comcast.net/~ericblake/eblake.gpg
Comment: Using GnuPG with Mozilla - http://enigmail.mozdev.org

iEYEARECAAYFAkgT8woACgkQ84KuGfSFAYCe/ACgzlnkWBaPhl+lTNAflqbzlWSK
BrcAn37vWsSgmIihj1Fqvgzs6DHZJCna
=/eUB
-----END PGP SIGNATURE-----

>From dc30711b753b14952eb670eea1807fbf0eab8a89 Mon Sep 17 00:00:00 2001
From: Eric Blake <address@hidden>
Date: Sat, 26 Apr 2008 21:26:10 -0600
Subject: [PATCH] Optimize and test memrchr.

* modules/memrchr (Depends-on): Add intprops.
* lib/memrchr.c (__memrchr): Avoid false positives in loop.
* modules/memrchr-tests: New file.
* tests/test-memrchr.c: New file.

Signed-off-by: Eric Blake <address@hidden>
---
 ChangeLog             |    9 ++
 lib/memrchr.c         |  198 +++++++++++++++++++++----------------------------
 modules/memrchr       |    1 +
 modules/memrchr-tests |   10 +++
 tests/test-memrchr.c  |   98 ++++++++++++++++++++++++
 5 files changed, 203 insertions(+), 113 deletions(-)
 create mode 100644 modules/memrchr-tests
 create mode 100644 tests/test-memrchr.c

diff --git a/ChangeLog b/ChangeLog
index b005118..d932986 100644
--- a/ChangeLog
+++ b/ChangeLog
@@ -1,3 +1,12 @@
+2008-04-26  Eric Blake  <address@hidden>
+       and Bruno Haible  <address@hidden>
+
+       Optimize and test memrchr.
+       * modules/memrchr (Depends-on): Add intprops.
+       * lib/memrchr.c (__memrchr): Avoid false positives in loop.
+       * modules/memrchr-tests: New file.
+       * tests/test-memrchr.c: New file.
+
 2008-04-26  Bruno Haible  <address@hidden>
 
        * tests/test-xstrtol.sh: Work around limitation of an old 'tr' program
diff --git a/lib/memrchr.c b/lib/memrchr.c
index 50b27c0..f4467ac 100644
--- a/lib/memrchr.c
+++ b/lib/memrchr.c
@@ -1,7 +1,7 @@
 /* memrchr -- find the last occurrence of a byte in a memory block
 
    Copyright (C) 1991, 1993, 1996, 1997, 1999, 2000, 2003, 2004, 2005,
-   2006, 2007 Free Software Foundation, Inc.
+   2006, 2007, 2008 Free Software Foundation, Inc.
 
    Based on strlen implementation by Torbjorn Granlund (address@hidden),
    with help from Dan Sahlin (address@hidden) and
@@ -32,6 +32,8 @@
 #include <string.h>
 #include <limits.h>
 
+#include "intprops.h"
+
 #undef __memrchr
 #undef memrchr
 
@@ -43,146 +45,116 @@
 void *
 __memrchr (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 longword;
+
   const unsigned char *char_ptr;
-  const unsigned long int *longword_ptr;
-  unsigned long int longword, magic_bits, charmask;
+  const longword *longword_ptr;
+  longword repeated_one;
+  longword repeated_c;
   unsigned reg_char c;
-  int i;
 
   c = (unsigned char) c_in;
 
-  /* Handle the last few characters by reading one character at a time.
+  /* Handle the last 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;
-       n > 0 && (size_t) char_ptr % sizeof longword != 0;
+       n > 0 && (size_t) char_ptr % sizeof (longword) != 0;
        --n)
     if (*--char_ptr == c)
       return (void *) char_ptr;
 
+  longword_ptr = (const longword *) char_ptr;
+
   /* All these elucidatory comments refer to 4-byte longwords,
      but the theory applies equally well to any size longwords.  */
 
-  longword_ptr = (const 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.  */
-
-  /* Set MAGIC_BITS to be this pattern of 1 and 0 bits.
-     Set CHARMASK to be a longword, each of whose bytes is C.  */
-
-  magic_bits = 0xfefefefe;
-  charmask = c | (c << 8);
-  charmask |= charmask << 16;
-#if 0xffffffffU < ULONG_MAX
-  magic_bits |= magic_bits << 32;
-  charmask |= charmask << 32;
-  if (8 < sizeof longword)
-    for (i = 64; i < sizeof longword * 8; i *= 2)
-      {
-       magic_bits |= magic_bits << i;
-       charmask |= charmask << i;
-      }
-#endif
-  magic_bits = (ULONG_MAX >> 1) & (magic_bits | 1);
-
-  /* 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.  */
-  while (n >= sizeof longword)
+  /* 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 < TYPE_MAXIMUM (longword))
     {
-      /* 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.
-
-        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.
-
-        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!
-
-        So it ignores everything except 128's, when they're aligned
-        properly.
-
-        3) But wait!  Aren't we looking for C, not zero?
-        Good point.  So what we do is XOR LONGWORD with a longword,
-        each of whose bytes is C.  This turns each byte that is C
-        into a zero.  */
-
-      longword = *--longword_ptr ^ charmask;
-
-      /* Add MAGIC_BITS to LONGWORD.  */
-      if ((((longword + magic_bits)
-
-           /* Set those bits that were unchanged by the addition.  */
-           ^ ~longword)
-
-          /* Look at only the hole bits.  If any of the hole bits
-             are unchanged, most likely one of the bytes was a
-             zero.  */
-          & ~magic_bits) != 0)
+      repeated_one |= repeated_one << 31 << 1;
+      repeated_c |= repeated_c << 31 << 1;
+      if (8 < sizeof (longword))
        {
-         /* Which of the bytes was C?  If none of them were, it was
-            a misfire; continue the search.  */
-
-         const unsigned char *cp = (const unsigned char *) longword_ptr;
-
-         if (8 < sizeof longword)
-           for (i = sizeof longword - 1; 8 <= i; i--)
-             if (cp[i] == c)
-               return (void *) &cp[i];
-         if (7 < sizeof longword && cp[7] == c)
-           return (void *) &cp[7];
-         if (6 < sizeof longword && cp[6] == c)
-           return (void *) &cp[6];
-         if (5 < sizeof longword && cp[5] == c)
-           return (void *) &cp[5];
-         if (4 < sizeof longword && cp[4] == c)
-           return (void *) &cp[4];
-         if (cp[3] == c)
-           return (void *) &cp[3];
-         if (cp[2] == c)
-           return (void *) &cp[2];
-         if (cp[1] == c)
-           return (void *) &cp[1];
-         if (cp[0] == c)
-           return (void *) cp;
+         size_t i;
+
+         for (i = 64; i < sizeof (longword) * 8; i *= 2)
+           {
+             repeated_one |= repeated_one << i;
+             repeated_c |= repeated_c << i;
+           }
        }
+    }
+
+  /* Instead of the traditional loop which tests each byte, 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 equal to c.  We first use an xor
+     with repeated_c.  This reduces the task to testing whether *any of the
+     four* bytes in longword1 is zero.
+
+     We compute tmp =
+       ((longword1 - repeated_one) & ~longword1) & (repeated_one << 7).
+     That is, we perform the following operations:
+       1. Subtract repeated_one.
+       2. & ~longword1.
+       3. & a mask consisting of 0x80 in every byte.
+     Consider what happens in each byte:
+       - If a byte of longword1 is zero, step 1 and 2 transform it into 0xff,
+         and step 3 transforms it into 0x80.  A carry can also be propagated
+         to more significant bytes.
+       - If a byte of longword1 is nonzero, let its lowest 1 bit be at
+         position k (0 <= k <= 7); so the lowest k bits are 0.  After step 1,
+         the byte ends in a single bit of value 0 and k bits of value 1.
+         After step 2, the result is just k bits of value 1: 2^k - 1.  After
+         step 3, the result is 0.  And no carry is produced.
+     So, if longword1 has only non-zero bytes, tmp is zero.
+     Whereas if longword1 has a zero byte, call j the position of the least
+     significant zero byte.  Then the result has a zero at positions 0, ...,
+     j-1 and a 0x80 at position j.  We cannot predict the result at the more
+     significant bytes (positions j+1..3), but it does not matter since we
+     already have a non-zero bit at position 8*j+7.
+
+     So, the test whether any byte in longword1 is zero is equivalent to
+     testing whether tmp is nonzero.  */
+
+  while (n >= sizeof (longword))
+    {
+      longword longword1 = *--longword_ptr ^ repeated_c;
 
-      n -= sizeof longword;
+      if ((((longword1 - repeated_one) & ~longword1)
+          & (repeated_one << 7)) != 0)
+       {
+         longword_ptr++;
+         break;
+       }
+      n -= sizeof (longword);
     }
 
   char_ptr = (const unsigned char *) longword_ptr;
 
+  /* At this point, we know that either n < sizeof (longword), or one of the
+     sizeof (longword) bytes starting at char_ptr is == c.  On little-endian
+     machines, we could determine the first such byte without any further
+     memory accesses, just by looking at the tmp result from the last loop
+     iteration.  But this does not work on big-endian machines.  Choose code
+     that works in both cases.  */
+
   while (n-- > 0)
     {
       if (*--char_ptr == c)
        return (void *) char_ptr;
     }
 
-  return 0;
+  return NULL;
 }
 #ifdef weak_alias
 weak_alias (__memrchr, memrchr)
diff --git a/modules/memrchr b/modules/memrchr
index 3a6e982..36a119d 100644
--- a/modules/memrchr
+++ b/modules/memrchr
@@ -7,6 +7,7 @@ m4/memrchr.m4
 
 Depends-on:
 extensions
+intprops
 string
 
 configure.ac:
diff --git a/modules/memrchr-tests b/modules/memrchr-tests
new file mode 100644
index 0000000..bad0515
--- /dev/null
+++ b/modules/memrchr-tests
@@ -0,0 +1,10 @@
+Files:
+tests/test-memrchr.c
+
+Depends-on:
+
+configure.ac:
+
+Makefile.am:
+TESTS += test-memrchr
+check_PROGRAMS += test-memrchr
diff --git a/tests/test-memrchr.c b/tests/test-memrchr.c
new file mode 100644
index 0000000..84f6c5b
--- /dev/null
+++ b/tests/test-memrchr.c
@@ -0,0 +1,98 @@
+/*
+ * Copyright (C) 2008 Free Software Foundation
+ * Written by Eric Blake and Bruno Haible
+ *
+ * This program is free software: you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation; either version 3 of the License, or
+ * (at your option) any later version.
+ *
+ * This program 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 General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program.  If not, see <http://www.gnu.org/licenses/>.  */
+
+#include <config.h>
+
+#include <string.h>
+
+#include <stdio.h>
+#include <stdlib.h>
+
+#define ASSERT(expr) \
+  do                                                                        \
+    {                                                                       \
+      if (!(expr))                                                          \
+       {                                                                    \
+         fprintf (stderr, "%s:%d: assertion failed\n", __FILE__, __LINE__); \
+         fflush (stderr);                                                   \
+         abort ();                                                          \
+       }                                                                    \
+    }                                                                       \
+  while (0)
+
+/* Calculating void * + int is not portable, so this wrapper converts
+   to char * to make the tests easier to write.  */
+#define MEMRCHR (char *) memrchr
+
+int
+main ()
+{
+  size_t n = 0x100000;
+  char *input = malloc (n);
+  ASSERT (input);
+
+  input[n - 1] = 'a';
+  input[n - 2] = 'b';
+  memset (input + n - 1026, 'c', 1024);
+  memset (input + 2, 'd', n - 1028);
+  input[1] = 'e';
+  input[0] = 'a';
+
+  /* Basic behavior tests.  */
+  ASSERT (MEMRCHR (input, 'a', n) == input + n - 1);
+
+  ASSERT (MEMRCHR (input, 'a', 0) == NULL);
+  ASSERT (MEMRCHR (NULL, 'a', 0) == NULL);
+
+  ASSERT (MEMRCHR (input, 'b', n) == input + n - 2);
+  ASSERT (MEMRCHR (input, 'c', n) == input + n - 3);
+  ASSERT (MEMRCHR (input, 'd', n) == input + n - 1027);
+
+  ASSERT (MEMRCHR (input, 'a', n - 1) == input);
+  ASSERT (MEMRCHR (input, 'e', n - 1) == input + 1);
+
+  ASSERT (MEMRCHR (input, 'f', n) == NULL);
+  ASSERT (MEMRCHR (input, '\0', n) == NULL);
+
+  /* Check that a very long haystack is handled quickly if the byte is
+     found near the end.  */
+  {
+    size_t repeat = 10000;
+    for (; repeat > 0; repeat--)
+      {
+       ASSERT (MEMRCHR (input, 'c', n) == input + n - 3);
+      }
+  }
+
+  /* Alignment tests.  */
+  {
+    int i, j;
+    for (i = 0; i < 32; i++)
+      {
+       for (j = 0; j < 256; j++)
+         input[i + j] = j;
+       for (j = 0; j < 256; j++)
+         {
+           ASSERT (MEMRCHR (input + i, j, 256) == input + i + j);
+         }
+      }
+  }
+
+  free (input);
+
+  return 0;
+}
-- 
1.5.5.1