string: Improve generic memchr

New algorithm read the first aligned address and mask off the
unwanted bytes (this strategy is similar to arch-specific
implementations used on powerpc, sparc, and sh).

The loop now read word-aligned address and check using the has_eq
macro.

Checked on x86_64-linux-gnu, i686-linux-gnu, powerpc-linux-gnu,
and powerpc64-linux-gnu by removing the arch-specific assembly
implementation and disabling multi-arch (it covers both LE and BE
for 64 and 32 bits).

Co-authored-by: Richard Henderson <richard.henderson@linaro.org>
Reviewed-by: Noah Goldstein <goldstein.w.n@gmail.com>

3 files changed, 56 insertions, 143 deletions

diff --git a/string/memchr.c b/string/memchr.c
index f800d47dce..4f497d6166 100644
--- a/string/memchr.c
+++ b/string/memchr.c
@@ -1,10 +1,6 @@
-/* Copyright (C) 1991-2023 Free Software Foundation, Inc.
+/* Scan memory for a character.  Generic version
+   Copyright (C) 1991-2023 Free Software Foundation, Inc.
    This file is part of the GNU C Library.
-   Based on strlen implementation by Torbjorn Granlund (tege@sics.se),
-   with help from Dan Sahlin (dan@sics.se) and
-   commentary by Jim Blandy (jimb@ai.mit.edu);
-   adaptation to memchr suggested by Dick Karpinski (dick@cca.ucsf.edu),
-   and implemented by Roland McGrath (roland@ai.mit.edu).
 
    The GNU C Library is free software; you can redistribute it and/or
    modify it under the terms of the GNU Lesser General Public
@@ -20,143 +16,73 @@
    License along with the GNU C Library; if not, see
    <https://www.gnu.org/licenses/>.  */
 
-#ifndef _LIBC
-# include <config.h>
-#endif
-
+#include <libc-pointer-arith.h>
+#include <string-fzb.h>
+#include <string-fzc.h>
+#include <string-fzi.h>
+#include <string-shift.h>
 #include <string.h>
 
-#include <stddef.h>
+#undef memchr
 
-#include <limits.h>
-
-#undef __memchr
-#ifdef _LIBC
-# undef memchr
+#ifdef MEMCHR
+# define __memchr MEMCHR
 #endif
 
-#ifndef weak_alias
-# define __memchr memchr
-#endif
-
-#ifndef MEMCHR
-# define MEMCHR __memchr
-#endif
+static __always_inline const char *
+sadd (uintptr_t x, uintptr_t y)
+{
+  return (const char *)(y > UINTPTR_MAX - x ? UINTPTR_MAX : x + y);
+}
 
 /* Search no more than N bytes of S for C.  */
 void *
-MEMCHR (void const *s, int c_in, size_t n)
+__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)
+  if (__glibc_unlikely (n == 0))
+    return NULL;
+
+  /* Read the first word, but munge it so that bytes before the array
+     will not match goal.  */
+  const op_t *word_ptr = PTR_ALIGN_DOWN (s, sizeof (op_t));
+  uintptr_t s_int = (uintptr_t) s;
+
+  op_t word = *word_ptr;
+  op_t repeated_c = repeat_bytes (c_in);
+  /* Compute the address of the last byte taking in consideration possible
+     overflow.  */
+  const char *lbyte = sadd (s_int, n - 1);
+  /* And also the address of the word containing the last byte. */
+  const op_t *lword = (const op_t *) PTR_ALIGN_DOWN (lbyte, sizeof (op_t));
+
+  find_t mask = shift_find (find_eq_all (word, repeated_c), s_int);
+  if (mask != 0)
     {
-      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;
-	    }
-	}
+      char *ret = (char *) s + index_first (mask);
+      return (ret <= lbyte) ? ret : NULL;
     }
+  if (word_ptr == lword)
+    return NULL;
 
-  /* 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))
+  word = *++word_ptr;
+  while (word_ptr != lword)
     {
-      longword longword1 = *longword_ptr ^ repeated_c;
-
-      if ((((longword1 - repeated_one) & ~longword1)
-	   & (repeated_one << 7)) != 0)
-	break;
-      longword_ptr++;
-      n -= sizeof (longword);
+      if (has_eq (word, repeated_c))
+	return (char *) word_ptr + index_first_eq (word, repeated_c);
+      word = *++word_ptr;
     }
 
-  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.  */
-
-  for (; n > 0; --n, ++char_ptr)
+  if (has_eq (word, repeated_c))
     {
-      if (*char_ptr == c)
-	return (void *) char_ptr;
+      /* We found a match, but it might be in a byte past the end of the
+	 array.  */
+      char *ret = (char *) word_ptr + index_first_eq (word, repeated_c);
+      if (ret <= lbyte)
+	return ret;
     }
-
   return NULL;
 }
-#ifdef weak_alias
+#ifndef MEMCHR
 weak_alias (__memchr, memchr)
-#endif
 libc_hidden_builtin_def (memchr)
+#endif
diff --git a/sysdeps/powerpc/powerpc32/power4/multiarch/memchr-ppc32.c b/sysdeps/powerpc/powerpc32/power4/multiarch/memchr-ppc32.c
index 39ff84f3f3..a78585650f 100644
--- a/sysdeps/powerpc/powerpc32/power4/multiarch/memchr-ppc32.c
+++ b/sysdeps/powerpc/powerpc32/power4/multiarch/memchr-ppc32.c
@@ -18,17 +18,11 @@
 
 #include <string.h>
 
-#define MEMCHR  __memchr_ppc
+extern __typeof (memchr) __memchr_ppc attribute_hidden;
 
-#undef weak_alias
-#define weak_alias(a, b)
+#define MEMCHR  __memchr_ppc
+#include <string/memchr.c>
 
 #ifdef SHARED
-# undef libc_hidden_builtin_def
-# define libc_hidden_builtin_def(name) \
-  __hidden_ver1(__memchr_ppc, __GI_memchr, __memchr_ppc);
+__hidden_ver1(__memchr_ppc, __GI_memchr, __memchr_ppc);
 #endif
-
-extern __typeof (memchr) __memchr_ppc attribute_hidden;
-
-#include <string/memchr.c>
diff --git a/sysdeps/powerpc/powerpc64/multiarch/memchr-ppc64.c b/sysdeps/powerpc/powerpc64/multiarch/memchr-ppc64.c
index 8097df709c..49ba5521fe 100644
--- a/sysdeps/powerpc/powerpc64/multiarch/memchr-ppc64.c
+++ b/sysdeps/powerpc/powerpc64/multiarch/memchr-ppc64.c
@@ -18,14 +18,7 @@
 
 #include <string.h>
 
-#define MEMCHR  __memchr_ppc
-
-#undef weak_alias
-#define weak_alias(a, b)
-
-# undef libc_hidden_builtin_def
-# define libc_hidden_builtin_def(name)
-
 extern __typeof (memchr) __memchr_ppc attribute_hidden;
 
+#define MEMCHR  __memchr_ppc
 #include <string/memchr.c>