Improve strcoll with strdiff.

This patch improves strcoll hot case by finding first byte that
mismatches. That is in likely case enough to determine comparison
result.

8 files changed, 71 insertions, 2 deletions

diff --git a/ChangeLog b/ChangeLog
index c998dd59b8..d78c626b25 100644
--- a/ChangeLog
+++ b/ChangeLog
@@ -1,3 +1,14 @@
+2015-05-12  Leonhard Holz <leonhard.holz@web.de>
+
+	* locale/categories.def: Define _NL_COLLATE_ENCODING_TYPE.
+	* locale/langinfo.h: Add _NL_COLLATE_ENCODING_TYPE to attribute list.
+	* locale/localeinfo.h: Add enum collation_encoding_type.
+	* locale/C-collate.c: Set _NL_COLLATE_ENCODING_TYPE to 8bit.
+	* programs/ld-collate.c (collate_output): Add encoding type info.
+	* string/strcoll_l.c (STRDIFF): New function.
+	* (STRCOLL): Use STRDIFF to skip over equal prefix.
+	* wcsmbs/wcscoll_l.c: Define STRDIFF.
+
 2015-05-11  Joseph Myers  <joseph@codesourcery.com>
 
 	[BZ #18397]
diff --git a/locale/C-collate.c b/locale/C-collate.c
index 06dfdfaad5..d7f3c550a5 100644
--- a/locale/C-collate.c
+++ b/locale/C-collate.c
@@ -144,6 +144,8 @@ const struct __locale_data _nl_C_LC_COLLATE attribute_hidden =
     /* _NL_COLLATE_COLLSEQWC */
     { .string = (const char *) collseqwc },
     /* _NL_COLLATE_CODESET */
-    { .string = _nl_C_codeset }
+    { .string = _nl_C_codeset },
+    /* _NL_COLLATE_ENCODING_TYPE */
+    { .word = __cet_8bit }
   }
 };
diff --git a/locale/categories.def b/locale/categories.def
index a8dda53007..045489d741 100644
--- a/locale/categories.def
+++ b/locale/categories.def
@@ -58,6 +58,7 @@ DEFINE_CATEGORY
   DEFINE_ELEMENT (_NL_COLLATE_COLLSEQMB,        "collate-collseqmb",        std, wstring)
   DEFINE_ELEMENT (_NL_COLLATE_COLLSEQWC,        "collate-collseqwc",        std, wstring)
   DEFINE_ELEMENT (_NL_COLLATE_CODESET,		"collate-codeset",	    std, string)
+  DEFINE_ELEMENT (_NL_COLLATE_ENCODING_TYPE,   "collate-encoding-type",    std, word)
   ), NO_POSTLOAD)
 
 
diff --git a/locale/langinfo.h b/locale/langinfo.h
index a565d9d120..ffc5c7f471 100644
--- a/locale/langinfo.h
+++ b/locale/langinfo.h
@@ -255,6 +255,7 @@ enum
   _NL_COLLATE_COLLSEQMB,
   _NL_COLLATE_COLLSEQWC,
   _NL_COLLATE_CODESET,
+  _NL_COLLATE_ENCODING_TYPE,
   _NL_NUM_LC_COLLATE,
 
   /* LC_CTYPE category: character classification.
diff --git a/locale/localeinfo.h b/locale/localeinfo.h
index 1d2ee00876..bdab9fe745 100644
--- a/locale/localeinfo.h
+++ b/locale/localeinfo.h
@@ -110,6 +110,14 @@ enum coll_sort_rule
   sort_mask
 };
 
+/* Collation encoding type.  */
+enum collation_encoding_type
+{
+  __cet_other,
+  __cet_8bit,
+  __cet_utf8
+};
+
 /* We can map the types of the entries into a few categories.  */
 enum value_type
 {
diff --git a/locale/programs/ld-collate.c b/locale/programs/ld-collate.c
index dc0fe30a13..a39a94f2cc 100644
--- a/locale/programs/ld-collate.c
+++ b/locale/programs/ld-collate.c
@@ -32,6 +32,7 @@
 #include "linereader.h"
 #include "locfile.h"
 #include "elem-hash.h"
+#include "../localeinfo.h"
 
 /* Uncomment the following line in the production version.  */
 /* #define NDEBUG 1 */
@@ -2130,6 +2131,8 @@ collate_output (struct localedef_t *locale, const struct charmap_t *charmap,
 	  /* The words have to be handled specially.  */
 	  if (idx == _NL_ITEM_INDEX (_NL_COLLATE_SYMB_HASH_SIZEMB))
 	    add_locale_uint32 (&file, 0);
+	  else if (idx == _NL_ITEM_INDEX (_NL_COLLATE_ENCODING_TYPE))
+	    add_locale_uint32 (&file, __cet_other);
 	  else
 	    add_locale_empty (&file);
 	}
@@ -2493,6 +2496,12 @@ collate_output (struct localedef_t *locale, const struct charmap_t *charmap,
   add_locale_raw_data (&file, collate->mbseqorder, 256);
   add_locale_collseq_table (&file, &collate->wcseqorder);
   add_locale_string (&file, charmap->code_set_name);
+  if (strcmp (charmap->code_set_name, "UTF-8") == 0)
+    add_locale_uint32 (&file, __cet_utf8);
+  else if (charmap->mb_cur_max == 1)
+    add_locale_uint32 (&file, __cet_8bit);
+  else
+    add_locale_uint32 (&file, __cet_other);
   write_locale_data (output_path, LC_COLLATE, "LC_COLLATE", &file);
 
   obstack_free (&weightpool, NULL);
diff --git a/string/strcoll_l.c b/string/strcoll_l.c
index 658d5b9b90..0fa005f0f9 100644
--- a/string/strcoll_l.c
+++ b/string/strcoll_l.c
@@ -29,6 +29,7 @@
 # define STRING_TYPE char
 # define USTRING_TYPE unsigned char
 # define STRCOLL __strcoll_l
+# define STRDIFF __strdiff
 # define STRCMP strcmp
 # define WEIGHT_H "../locale/weight.h"
 # define SUFFIX	MB
@@ -41,6 +42,20 @@
 #include "../locale/localeinfo.h"
 #include WEIGHT_H
 
+#define MASK_UTF8_7BIT  (1 << 7)
+#define MASK_UTF8_START (3 << 6)
+
+size_t
+STRDIFF (const STRING_TYPE *s, const STRING_TYPE *t)
+{
+  size_t n;
+
+  for (n = 0; *s != '\0' && *s++ == *t++; ++n)
+    continue;
+
+  return n;
+}
+
 /* Track status while looking for sequences in a string.  */
 typedef struct
 {
@@ -255,9 +270,29 @@ STRCOLL (const STRING_TYPE *s1, const STRING_TYPE *s2, __locale_t l)
   const USTRING_TYPE *extra;
   const int32_t *indirect;
 
+  /* In case there is no locale specific sort order (C / POSIX).  */
   if (nrules == 0)
     return STRCMP (s1, s2);
 
+  /* Fast forward to the position of the first difference.  Needs to be
+     encoding aware as the byte-by-byte comparison can stop in the middle
+     of a char sequence for multibyte encodings like UTF-8.  */
+  uint_fast32_t encoding =
+    current->values[_NL_ITEM_INDEX (_NL_COLLATE_ENCODING_TYPE)].word;
+  if (encoding != __cet_other)
+    {
+      size_t diff = STRDIFF (s1, s2);
+      if (diff > 0)
+	{
+	  if (encoding == __cet_utf8 && (*(s1 + diff) & MASK_UTF8_7BIT) != 0)
+	    do
+	      diff--;
+	    while (diff > 0 && (*(s1 + diff) & MASK_UTF8_START) != MASK_UTF8_START);
+	  s1 += diff;
+	  s2 += diff;
+	}
+    }
+
   /* Catch empty strings.  */
   if (__glibc_unlikely (*s1 == '\0') || __glibc_unlikely (*s2 == '\0'))
     return (*s1 != '\0') - (*s2 != '\0');
@@ -321,7 +356,8 @@ STRCOLL (const STRING_TYPE *s1, const STRING_TYPE *s2, __locale_t l)
 		     byte-level comparison to ensure that we don't waste time
 		     going through multiple passes for totally equal strings
 		     before proceeding to subsequent passes.  */
-		  if (pass == 0 && STRCMP (s1, s2) == 0)
+		  if (pass == 0 && encoding == __cet_other &&
+		      STRCMP (s1, s2) == 0)
 		    return result;
 		  else
 		    break;
diff --git a/wcsmbs/wcscoll_l.c b/wcsmbs/wcscoll_l.c
index 106ec93511..9f60cee9ea 100644
--- a/wcsmbs/wcscoll_l.c
+++ b/wcsmbs/wcscoll_l.c
@@ -23,6 +23,7 @@
 #define STRING_TYPE wchar_t
 #define USTRING_TYPE wint_t
 #define STRCOLL __wcscoll_l
+#define STRDIFF __wcsdiff
 #define STRCMP wcscmp
 #define WEIGHT_H "../locale/weightwc.h"
 #define SUFFIX	WC