Improve strstr performance

Improve strstr performance.  Strstr tends to be slow because it uses
many calls to memchr and a slow byte loop to scan for the next match.
Performance is significantly improved by using strnlen on larger blocks
and using strchr to search for the next matching character.  strcasestr
can also use strnlen to scan ahead, and memmem can use memchr to check
for the next match.

On the GLIBC bench tests the performance gains on Cortex-A72 are:
strstr: +25%
strcasestr: +4.3%
memmem: +18%

On a 256KB dataset strstr performance improves by 67%, strcasestr by 47%.

    Reviewd-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>

ChangeLog

@@ -1,3 +1,16 @@
+2018-07-15  Wilco Dijkstra  <wdijkstr@arm.com>
+
+	* benchtests/bench-strcasestr.c: Rename __strnlen to strnlen.
+	* benchtests/bench-strstr.c: Likewise.
+	* string/memmem.c (FASTSEARCH): Define.
+	* string/str-two-way.h (two_way_short_needle): Minor cleanups.
+	Add support for FASTSEARCH.
+	* string/strcasestr.c (AVAILABLE): Use read-ahead __strnlen.
+	* string/strstr.c (AVAILABLE): Use read-ahead __strnlen.
+	(FASTSEARCH): Define.
+	* string/test-strcasestr.c: Rename __strnlen to strnlen.
+	* string/test-strstr.c: Likewise.
+
 2018-07-15  H.J. Lu  <hongjiu.lu@intel.com>
 
 	* sysdeps/unix/sysv/linux/x86_64/____longjmp_chk.S: Undef

benchtests/bench-strcasestr.c

@@ -24,6 +24,7 @@
 #define STRCASESTR simple_strcasestr
 #define NO_ALIAS
 #define __strncasecmp strncasecmp
+#define __strnlen strnlen
 #include "../string/strcasestr.c"
 
 

benchtests/bench-strstr.c

@@ -23,6 +23,7 @@
 
 #define STRSTR simple_strstr
 #define libc_hidden_builtin_def(X)
+#define __strnlen strnlen
 #include "../string/strstr.c"
 
 

string/memmem.c

@@ -31,6 +31,7 @@
 
 #define RETURN_TYPE void *
 #define AVAILABLE(h, h_l, j, n_l) ((j) <= (h_l) - (n_l))
+#define FASTSEARCH(S,C,N) (void*) memchr ((void *)(S), (C), (N))
 #include "str-two-way.h"
 
 #undef memmem

string/str-two-way.h

@@ -281,50 +281,50 @@ two_way_short_needle (const unsigned char *haystack, size_t haystack_len,
     }
   else
     {
-      const unsigned char *phaystack = &haystack[suffix];
+      const unsigned char *phaystack;
       /* The comparison always starts from needle[suffix], so cache it
 	 and use an optimized first-character loop.  */
       unsigned char needle_suffix = CANON_ELEMENT (needle[suffix]);
 
-#if CHECK_EOL
-      /* We start matching from the SUFFIX'th element, so make sure we
-	 don't hit '\0' before that.  */
-      if (haystack_len < suffix + 1
-	  && !AVAILABLE (haystack, haystack_len, 0, suffix + 1))
-	return NULL;
-#endif
-
       /* The two halves of needle are distinct; no extra memory is
 	 required, and any mismatch results in a maximal shift.  */
       period = MAX (suffix, needle_len - suffix) + 1;
       j = 0;
-      while (1
-#if !CHECK_EOL
-	     && AVAILABLE (haystack, haystack_len, j, needle_len)
-#endif
-	     )
+      while (AVAILABLE (haystack, haystack_len, j, needle_len))
 	{
 	  unsigned char haystack_char;
 	  const unsigned char *pneedle;
 
-	  /* TODO: The first-character loop can be sped up by adapting
-	     longword-at-a-time implementation of memchr/strchr.  */
-	  if (needle_suffix
+	  phaystack = &haystack[suffix + j];
+
+#ifdef FASTSEARCH
+	  if (*phaystack++ != needle_suffix)
+	    {
+	      phaystack = FASTSEARCH (phaystack, needle_suffix,
+				      haystack_len - needle_len - j);
+	      if (phaystack == NULL)
+		goto ret0;
+	      j = phaystack - &haystack[suffix];
+	      phaystack++;
+	    }
+#else
+	  while (needle_suffix
 	      != (haystack_char = CANON_ELEMENT (*phaystack++)))
 	    {
 	      RET0_IF_0 (haystack_char);
-#if !CHECK_EOL
+# if !CHECK_EOL
 	      ++j;
-#endif
-	      continue;
+	      if (!AVAILABLE (haystack, haystack_len, j, needle_len))
+		goto ret0;
+# endif
 	    }
 
-#if CHECK_EOL
+# if CHECK_EOL
 	  /* Calculate J if it wasn't kept up-to-date in the first-character
 	     loop.  */
 	  j = phaystack - &haystack[suffix] - 1;
+# endif
 #endif
-
 	  /* Scan for matches in right half.  */
 	  i = suffix + 1;
 	  pneedle = &needle[i];
@@ -338,6 +338,11 @@ two_way_short_needle (const unsigned char *haystack, size_t haystack_len,
 		}
 	      ++i;
 	    }
+#if CHECK_EOL
+	  /* Update minimal length of haystack.  */
+	  if (phaystack > haystack + haystack_len)
+	    haystack_len = phaystack - haystack;
+#endif
 	  if (needle_len <= i)
 	    {
 	      /* Scan for matches in left half.  */
@@ -360,13 +365,6 @@ two_way_short_needle (const unsigned char *haystack, size_t haystack_len,
 	    }
 	  else
 	    j += i - suffix + 1;
-
-#if CHECK_EOL
-	  if (!AVAILABLE (haystack, haystack_len, j, needle_len))
-	    break;
-#endif
-
-	  phaystack = &haystack[suffix + j];
 	}
     }
  ret0: __attribute__ ((unused))

string/strcasestr.c

@@ -37,8 +37,8 @@
 /* Two-Way algorithm.  */
 #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)))
+  (((j) + (n_l) <= (h_l)) || ((h_l) += __strnlen ((void*)((h) + (h_l)), 512), \
+			      (j) + (n_l) <= (h_l)))
 #define CHECK_EOL (1)
 #define RET0_IF_0(a) if (!a) goto ret0
 #define CANON_ELEMENT(c) TOLOWER (c)

string/strstr.c

@@ -33,10 +33,11 @@
 
 #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)))
+  (((j) + (n_l) <= (h_l)) || ((h_l) += __strnlen ((void*)((h) + (h_l)), 512), \
+			      (j) + (n_l) <= (h_l)))
 #define CHECK_EOL (1)
 #define RET0_IF_0(a) if (!a) goto ret0
+#define FASTSEARCH(S,C,N) (void*) strchr ((void*)(S), (C))
 #include "str-two-way.h"
 
 #undef strstr

string/test-strcasestr.c

@@ -25,6 +25,7 @@
 #define STRCASESTR simple_strcasestr
 #define NO_ALIAS
 #define __strncasecmp strncasecmp
+#define __strnlen strnlen
 #include "strcasestr.c"
 
 

string/test-strstr.c

@@ -24,6 +24,7 @@
 
 #define STRSTR simple_strstr
 #define libc_hidden_builtin_def(arg) /* nothing */
+#define __strnlen strnlen
 #include "strstr.c"