[v3,06/18] string: Improve generic memchr
Commit Message
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(-)
@@ -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)