diff mbox series

[v3,06/18] string: Improve generic memchr

Message ID 1515588482-15744-7-git-send-email-adhemerval.zanella@linaro.org
State New
Headers show
Series Improve generic string routines | expand

Commit Message

Adhemerval Zanella Jan. 10, 2018, 12:47 p.m. UTC
From: Richard Henderson <rth@twiddle.net>


New algorithm have the following key differences:

  - Reads first word unaligned and use string-maskoff function to
    remove unwanted data.  This strategy follow assemble optimized
    ones for aarch64, powerpc and tile.

  - Use string-fz{b,i} and string-opthr functions.

Checked on x86_64-linux-gnu, i686-linux-gnu, sparc64-linux-gnu,
and sparcv9-linux-gnu by removing the arch-specific assembly
implementation and disabling multi-arch (it covers both LE and BE
for 64 and 32 bits).

	[BZ #5806]
	* string/memchr.c: Use string-fzb.h, string-fzi.h, string-opthr.h.
---
 string/memchr.c | 157 +++++++++++++++-----------------------------------------
 1 file changed, 40 insertions(+), 117 deletions(-)

-- 
2.7.4
diff mbox series

Patch

diff --git a/string/memchr.c b/string/memchr.c
index c4e21b8..ae3fd93 100644
--- a/string/memchr.c
+++ b/string/memchr.c
@@ -20,24 +20,16 @@ 
    License along with the GNU C Library; if not, see
    <http://www.gnu.org/licenses/>.  */
 
-#ifndef _LIBC
-# include <config.h>
-#endif
-
 #include <string.h>
-
 #include <stddef.h>
+#include <stdint.h>
+#include <string-fza.h>
+#include <string-fzb.h>
+#include <string-fzi.h>
+#include <string-maskoff.h>
+#include <string-opthr.h>
 
-#include <limits.h>
-
-#undef __memchr
-#ifdef _LIBC
-# undef memchr
-#endif
-
-#ifndef weak_alias
-# define __memchr memchr
-#endif
+#undef memchr
 
 #ifndef MEMCHR
 # define MEMCHR __memchr
@@ -47,116 +39,47 @@ 
 void *
 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;
-
-  const unsigned char *char_ptr;
-  const longword *longword_ptr;
-  longword repeated_one;
-  longword repeated_c;
-  unsigned char c;
-
-  c = (unsigned char) c_in;
-
-  /* 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;
-
-  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.  */
-
-  /* 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)
-    {
-      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;
-	    }
-	}
-    }
+  const op_t *word_ptr, *lword;
+  op_t repeated_c, before_mask, word;
+  const char *lbyte;
+  char *ret;
+  uintptr_t s_int;
 
-  /* 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;
 
-      if ((((longword1 - repeated_one) & ~longword1)
-	   & (repeated_one << 7)) != 0)
-	break;
-      longword_ptr++;
-      n -= sizeof (longword);
-    }
+  if (__glibc_unlikely (n == 0))
+    return NULL;
+
+  s_int = (uintptr_t) s;
+  word_ptr = (const op_t*) (s_int & -sizeof (op_t));
 
-  char_ptr = (const unsigned char *) longword_ptr;
+  /* Set up a word, each of whose bytes is C.  */
+  repeated_c = repeat_bytes (c_in);
+  before_mask = create_mask (s_int);
 
-  /* 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.  */
+  /* Compute the address of the last byte taking in consideration possible
+     overflow.  */
+  uintptr_t lbyte_int = s_int + n - 1;
+  lbyte_int |= -(lbyte_int < s_int);
+  lbyte = (const char *) lbyte_int;
 
-  for (; n > 0; --n, ++char_ptr)
+  /* Compute the address of the word containing the last byte. */
+  lword = (const op_t *) ((uintptr_t) lbyte & -sizeof (op_t));
+
+  /* Read the first word, but munge it so that bytes before the array
+     will not match goal.  */
+  word = (*word_ptr | before_mask) ^ (repeated_c & before_mask);
+
+  while (has_eq (word, repeated_c) == 0)
     {
-      if (*char_ptr == c)
-	return (void *) char_ptr;
+      if (word_ptr == lword)
+	return NULL;
+      word = *++word_ptr;
     }
 
-  return NULL;
+  /* We found a match, but it might be in a byte past the end
+     of the array.  */
+  ret = (char *) word_ptr + index_first_eq (word, repeated_c);
+  return (ret <= lbyte) ? ret : NULL;
 }
-#ifdef weak_alias
 weak_alias (__memchr, memchr)
-#endif
 libc_hidden_builtin_def (memchr)