From: Bruno Haible Date: Sun, 11 Feb 2007 21:33:27 +0000 (+0000) Subject: Ensure O(n) worst-case complexity of mbscasestr. X-Git-Tag: cvs-readonly~1083 X-Git-Url: http://erislabs.org.uk/gitweb/?a=commitdiff_plain;h=35aa104af1bc1bd8263bde7c4202406845403f83;p=gnulib.git Ensure O(n) worst-case complexity of mbscasestr. --- diff --git a/ChangeLog b/ChangeLog index 1f45bd193..a02adbbf8 100644 --- a/ChangeLog +++ b/ChangeLog @@ -1,5 +1,13 @@ 2007-02-11 Bruno Haible + Ensure O(n) worst-case complexity of mbscasestr. + * lib/mbscasestr.c: Include stdbool.h. + (knuth_morris_pratt_unibyte, knuth_morris_pratt_multibyte): New + functions. + (mbscasestr): Add some bookkeeping. Invoke knuth_morris_pratt_* when + the bookkeeping indicates that it's worth it. + * modules/mbscasestr (Depends-on): Add stdbool, mbslen, strnlen. + * modules/mbscasestr-tests: New file. * tests/test-mbscasestr1.c: New file. * tests/test-mbscasestr2.sh: New file. diff --git a/lib/mbscasestr.c b/lib/mbscasestr.c index cf23cbaa4..838120cba 100644 --- a/lib/mbscasestr.c +++ b/lib/mbscasestr.c @@ -22,6 +22,7 @@ #include #include +#include #include /* for NULL, in case a nonstandard string.h lacks it */ #if HAVE_MBRTOWC @@ -30,6 +31,217 @@ #define TOLOWER(Ch) (isupper (Ch) ? tolower (Ch) : (Ch)) +/* Knuth-Morris-Pratt algorithm. + See http://en.wikipedia.org/wiki/Knuth-Morris-Pratt_algorithm + Return a boolean indicating success. */ + +static bool +knuth_morris_pratt_unibyte (const char *haystack, const char *needle, + const char **resultp) +{ + size_t m = strlen (needle); + + /* Allocate the table. */ + size_t *table = (size_t *) malloc (m * sizeof (size_t)); + if (table == NULL) + return false; + /* Fill the table. + For 0 < i < m: + 0 < table[i] <= i is defined such that + rhaystack[0..i-1] == needle[0..i-1] and rhaystack[i] != needle[i] + implies + forall 0 <= x < table[i]: rhaystack[x..x+m-1] != needle[0..m-1], + and table[i] is as large as possible with this property. + table[0] remains uninitialized. */ + { + size_t i, j; + + table[1] = 1; + j = 0; + for (i = 2; i < m; i++) + { + unsigned char b = TOLOWER ((unsigned char) needle[i - 1]); + + for (;;) + { + if (b == TOLOWER ((unsigned char) needle[j])) + { + table[i] = i - ++j; + break; + } + if (j == 0) + { + table[i] = i; + break; + } + j = j - table[j]; + } + } + } + + /* Search, using the table to accelerate the processing. */ + { + size_t j; + const char *rhaystack; + const char *phaystack; + + *resultp = NULL; + j = 0; + rhaystack = haystack; + phaystack = haystack; + /* Invariant: phaystack = rhaystack + j. */ + while (*phaystack != '\0') + if (TOLOWER ((unsigned char) needle[j]) + == TOLOWER ((unsigned char) *phaystack)) + { + j++; + phaystack++; + if (j == m) + { + /* The entire needle has been found. */ + *resultp = rhaystack; + break; + } + } + else if (j > 0) + { + /* Found a match of needle[0..j-1], mismatch at needle[j]. */ + rhaystack += table[j]; + j -= table[j]; + } + else + { + /* Found a mismatch at needle[0] already. */ + rhaystack++; + phaystack++; + } + } + + free (table); + return true; +} + +#if HAVE_MBRTOWC +static bool +knuth_morris_pratt_multibyte (const char *haystack, const char *needle, + const char **resultp) +{ + size_t m = mbslen (needle); + mbchar_t *needle_mbchars; + size_t *table; + + /* Allocate room for needle_mbchars and the table. */ + char *memory = (char *) malloc (m * (sizeof (mbchar_t) + sizeof (size_t))); + if (memory == NULL) + return false; + needle_mbchars = (mbchar_t *) memory; + table = (size_t *) (memory + m * sizeof (mbchar_t)); + + /* Fill needle_mbchars. */ + { + mbui_iterator_t iter; + size_t j; + + j = 0; + for (mbui_init (iter, needle); mbui_avail (iter); mbui_advance (iter), j++) + { + mb_copy (&needle_mbchars[j], &mbui_cur (iter)); + if (needle_mbchars[j].wc_valid) + needle_mbchars[j].wc = towlower (needle_mbchars[j].wc); + } + } + + /* Fill the table. + For 0 < i < m: + 0 < table[i] <= i is defined such that + rhaystack[0..i-1] == needle[0..i-1] and rhaystack[i] != needle[i] + implies + forall 0 <= x < table[i]: rhaystack[x..x+m-1] != needle[0..m-1], + and table[i] is as large as possible with this property. + table[0] remains uninitialized. */ + { + size_t i, j; + + table[1] = 1; + j = 0; + for (i = 2; i < m; i++) + { + mbchar_t *b = &needle_mbchars[i - 1]; + + for (;;) + { + if (mb_equal (*b, needle_mbchars[j])) + { + table[i] = i - ++j; + break; + } + if (j == 0) + { + table[i] = i; + break; + } + j = j - table[j]; + } + } + } + + /* Search, using the table to accelerate the processing. */ + { + size_t j; + mbui_iterator_t rhaystack; + mbui_iterator_t phaystack; + + *resultp = NULL; + j = 0; + mbui_init (rhaystack, haystack); + mbui_init (phaystack, haystack); + /* Invariant: phaystack = rhaystack + j. */ + while (mbui_avail (phaystack)) + { + mbchar_t c; + + mb_copy (&c, &mbui_cur (phaystack)); + if (c.wc_valid) + c.wc = towlower (c.wc); + if (mb_equal (needle_mbchars[j], c)) + { + j++; + mbui_advance (phaystack); + if (j == m) + { + /* The entire needle has been found. */ + *resultp = mbui_cur_ptr (rhaystack); + break; + } + } + else if (j > 0) + { + /* Found a match of needle[0..j-1], mismatch at needle[j]. */ + size_t count = table[j]; + j -= count; + for (; count > 0; count--) + { + if (!mbui_avail (rhaystack)) + abort (); + mbui_advance (rhaystack); + } + } + else + { + /* Found a mismatch at needle[0] already. */ + if (!mbui_avail (rhaystack)) + abort (); + mbui_advance (rhaystack); + mbui_advance (phaystack); + } + } + } + + free (memory); + return true; +} +#endif + /* Find the first occurrence of the character string NEEDLE in the character string HAYSTACK, using case-insensitive comparison. Note: This function may, in multibyte locales, return success even if @@ -50,9 +262,31 @@ mbscasestr (const char *haystack, const char *needle) mbui_init (iter_needle, needle); if (mbui_avail (iter_needle)) { + /* Minimizing the worst-case complexity: + Let n = mbslen(haystack), m = mbslen(needle). + The naïve algorithm is O(n*m) worst-case. + The Knuth-Morris-Pratt algorithm is O(n) worst-case but it needs a + memory allocation. + To achieve linear complexity and yet amortize the cost of the + memory allocation, we activate the Knuth-Morris-Pratt algorithm + only once the naïve algorithm has already run for some time; more + precisely, when + - the outer loop count is >= 10, + - the average number of comparisons per outer loop is >= 5, + - the total number of comparisons is >= m. + But we try it only once. If the memory allocation attempt failed, + we don't retry it. */ + bool try_kmp = true; + size_t outer_loop_count = 0; + size_t comparison_count = 0; + size_t last_ccount = 0; /* last comparison count */ + mbui_iterator_t iter_needle_last_ccount; /* = needle + last_ccount */ + mbchar_t b; mbui_iterator_t iter_haystack; + mbui_init (iter_needle_last_ccount, needle); + mb_copy (&b, &mbui_cur (iter_needle)); if (b.wc_valid) b.wc = towlower (b.wc); @@ -66,6 +300,35 @@ mbscasestr (const char *haystack, const char *needle) /* No match. */ return NULL; + /* See whether it's advisable to use an asymptotically faster + algorithm. */ + if (try_kmp + && outer_loop_count >= 10 + && comparison_count >= 5 * outer_loop_count) + { + /* See if needle + comparison_count now reaches the end of + needle. */ + size_t count = comparison_count - last_ccount; + for (; + count > 0 && mbui_avail (iter_needle_last_ccount); + count--) + mbui_advance (iter_needle_last_ccount); + last_ccount = comparison_count; + if (!mbui_avail (iter_needle_last_ccount)) + { + /* Try the Knuth-Morris-Pratt algorithm. */ + const char *result; + bool success = + knuth_morris_pratt_multibyte (haystack, needle, + &result); + if (success) + return (char *) result; + try_kmp = false; + } + } + + outer_loop_count++; + comparison_count++; mb_copy (&c, &mbui_cur (iter_haystack)); if (c.wc_valid) c.wc = towlower (c.wc); @@ -91,6 +354,7 @@ mbscasestr (const char *haystack, const char *needle) if (!mbui_avail (rhaystack)) /* No match. */ return NULL; + comparison_count++; if (!mb_caseequal (mbui_cur (rhaystack), mbui_cur (rneedle))) /* Nothing in this round. */ @@ -107,6 +371,26 @@ mbscasestr (const char *haystack, const char *needle) { if (*needle != '\0') { + /* Minimizing the worst-case complexity: + Let n = strlen(haystack), m = strlen(needle). + The naïve algorithm is O(n*m) worst-case. + The Knuth-Morris-Pratt algorithm is O(n) worst-case but it needs a + memory allocation. + To achieve linear complexity and yet amortize the cost of the + memory allocation, we activate the Knuth-Morris-Pratt algorithm + only once the naïve algorithm has already run for some time; more + precisely, when + - the outer loop count is >= 10, + - the average number of comparisons per outer loop is >= 5, + - the total number of comparisons is >= m. + But we try it only once. If the memory allocation attempt failed, + we don't retry it. */ + bool try_kmp = true; + size_t outer_loop_count = 0; + size_t comparison_count = 0; + size_t last_ccount = 0; /* last comparison count */ + const char *needle_last_ccount = needle; /* = needle + last_ccount */ + /* Speed up the following searches of needle by caching its first character. */ unsigned char b = TOLOWER ((unsigned char) *needle); @@ -117,6 +401,39 @@ mbscasestr (const char *haystack, const char *needle) if (*haystack == '\0') /* No match. */ return NULL; + + /* See whether it's advisable to use an asymptotically faster + algorithm. */ + if (try_kmp + && outer_loop_count >= 10 + && comparison_count >= 5 * outer_loop_count) + { + /* See if needle + comparison_count now reaches the end of + needle. */ + if (needle_last_ccount != NULL) + { + needle_last_ccount += + strnlen (needle_last_ccount, + comparison_count - last_ccount); + if (*needle_last_ccount == '\0') + needle_last_ccount = NULL; + last_ccount = comparison_count; + } + if (needle_last_ccount == NULL) + { + /* Try the Knuth-Morris-Pratt algorithm. */ + const char *result; + bool success = + knuth_morris_pratt_unibyte (haystack, needle - 1, + &result); + if (success) + return (char *) result; + try_kmp = false; + } + } + + outer_loop_count++; + comparison_count++; if (TOLOWER ((unsigned char) *haystack) == b) /* The first character matches. */ { @@ -131,6 +448,7 @@ mbscasestr (const char *haystack, const char *needle) if (*rhaystack == '\0') /* No match. */ return NULL; + comparison_count++; if (TOLOWER ((unsigned char) *rhaystack) != TOLOWER ((unsigned char) *rneedle)) /* Nothing in this round. */ diff --git a/modules/mbscasestr b/modules/mbscasestr index ec801a7d1..5a6bda95b 100644 --- a/modules/mbscasestr +++ b/modules/mbscasestr @@ -8,7 +8,10 @@ m4/mbrtowc.m4 Depends-on: mbuiter +stdbool string +mbslen +strnlen configure.ac: gl_FUNC_MBSCASESTR