#include "reentr.h"
+/* If the environment says to, we can output debugging information during
+ * initialization. This is done before option parsing, and before any thread
+ * creation, so can be a file-level static */
+#ifdef DEBUGGING
+static bool debug_initialization = FALSE;
+#endif
+
#ifdef USE_LOCALE
/*
else
PL_numeric_radix_sv = NULL;
- DEBUG_L(PerlIO_printf(Perl_debug_log, "Locale radix is '%s', ?UTF-8=%d\n",
+#ifdef DEBUGGING
+ if (DEBUG_L_TEST || debug_initialization) {
+ PerlIO_printf(Perl_debug_log, "Locale radix is '%s', ?UTF-8=%d\n",
(PL_numeric_radix_sv)
? SvPVX(PL_numeric_radix_sv)
: "NULL",
(PL_numeric_radix_sv)
? cBOOL(SvUTF8(PL_numeric_radix_sv))
- : 0));
+ : 0);
+ }
+#endif
# endif /* HAS_LOCALECONV */
#endif /* USE_LOCALE_NUMERIC */
PL_numeric_standard = TRUE;
PL_numeric_local = isNAME_C_OR_POSIX(PL_numeric_name);
set_numeric_radix();
- DEBUG_L(PerlIO_printf(Perl_debug_log,
- "Underlying LC_NUMERIC locale now is C\n"));
+#ifdef DEBUGGING
+ if (DEBUG_L_TEST || debug_initialization) {
+ PerlIO_printf(Perl_debug_log,
+ "Underlying LC_NUMERIC locale now is C\n");
+ }
+#endif
#endif /* USE_LOCALE_NUMERIC */
}
PL_numeric_standard = isNAME_C_OR_POSIX(PL_numeric_name);
PL_numeric_local = TRUE;
set_numeric_radix();
- DEBUG_L(PerlIO_printf(Perl_debug_log,
+#ifdef DEBUGGING
+ if (DEBUG_L_TEST || debug_initialization) {
+ PerlIO_printf(Perl_debug_log,
"Underlying LC_NUMERIC locale now is %s\n",
- PL_numeric_name));
+ PL_numeric_name);
+ }
+#endif
#endif /* USE_LOCALE_NUMERIC */
}
PL_collxfrm_mult = 2;
PL_in_utf8_COLLATE_locale = FALSE;
*PL_strxfrm_min_char = '\0';
+ PL_strxfrm_max_cp = 0;
return;
}
PL_in_utf8_COLLATE_locale = _is_cur_LC_category_utf8(LC_COLLATE);
*PL_strxfrm_min_char = '\0';
+ PL_strxfrm_max_cp = 0;
/* A locale collation definition includes primary, secondary, tertiary,
* etc. weights for each character. To sort, the primary weights are
* transformations. */
{
- /* 2: at most so many chars ('a', 'b'). */
- /* 50: surely no system expands a char more. */
-#define XFRMBUFSIZE (2 * 50)
- char xbuf[XFRMBUFSIZE];
- const Size_t fa = strxfrm(xbuf, "a", XFRMBUFSIZE);
- const Size_t fb = strxfrm(xbuf, "ab", XFRMBUFSIZE);
- const SSize_t mult = fb - fa;
- if (mult < 1 && !(fa == 0 && fb == 0))
- Perl_croak(aTHX_ "panic: strxfrm() gets absurd - a => %"UVuf", ab => %"UVuf,
- (UV) fa, (UV) fb);
- PL_collxfrm_base = (fa > (Size_t)mult) ? (fa - mult) : 0;
- PL_collxfrm_mult = mult;
+ /* We use the string below to find how long the tranformation of it
+ * is. Almost all locales are supersets of ASCII, or at least the
+ * ASCII letters. We use all of them, half upper half lower,
+ * because if we used fewer, we might hit just the ones that are
+ * outliers in a particular locale. Most of the strings being
+ * collated will contain a preponderance of letters, and even if
+ * they are above-ASCII, they are likely to have the same number of
+ * weight levels as the ASCII ones. It turns out that digits tend
+ * to have fewer levels, and some punctuation has more, but those
+ * are relatively sparse in text, and khw believes this gives a
+ * reasonable result, but it could be changed if experience so
+ * dictates. */
+ const char longer[] = "ABCDEFGHIJKLMnopqrstuvwxyz";
+ char * x_longer; /* Transformed 'longer' */
+ Size_t x_len_longer; /* Length of 'x_longer' */
+
+ char * x_shorter; /* We also transform a substring of 'longer' */
+ Size_t x_len_shorter;
+
+ /* _mem_collxfrm() is used get the transformation (though here we
+ * are interested only in its length). It is used because it has
+ * the intelligence to handle all cases, but to work, it needs some
+ * values of 'm' and 'b' to get it started. For the purposes of
+ * this calculation we use a very conservative estimate of 'm' and
+ * 'b'. This assumes a weight can be multiple bytes, enough to
+ * hold any UV on the platform, and there are 5 levels, 4 weight
+ * bytes, and a trailing NUL. */
+ PL_collxfrm_base = 5;
+ PL_collxfrm_mult = 5 * sizeof(UV);
+
+ /* Find out how long the transformation really is */
+ x_longer = _mem_collxfrm(longer,
+ sizeof(longer) - 1,
+ &x_len_longer,
+
+ /* We avoid converting to UTF-8 in the
+ * called function by telling it the
+ * string is in UTF-8 if the locale is a
+ * UTF-8 one. Since the string passed
+ * here is invariant under UTF-8, we can
+ * claim it's UTF-8 even though it isn't.
+ * */
+ PL_in_utf8_COLLATE_locale);
+ Safefree(x_longer);
+
+ /* Find out how long the transformation of a substring of 'longer'
+ * is. Together the lengths of these transformations are
+ * sufficient to calculate 'm' and 'b'. The substring is all of
+ * 'longer' except the first character. This minimizes the chances
+ * of being swayed by outliers */
+ x_shorter = _mem_collxfrm(longer + 1,
+ sizeof(longer) - 2,
+ &x_len_shorter,
+ PL_in_utf8_COLLATE_locale);
+ Safefree(x_shorter);
+
+ /* If the results are nonsensical for this simple test, the whole
+ * locale definition is suspect. Mark it so that locale collation
+ * is not active at all for it. XXX Should we warn? */
+ if ( x_len_shorter == 0
+ || x_len_longer == 0
+ || x_len_shorter >= x_len_longer)
+ {
+ PL_collxfrm_mult = 0;
+ PL_collxfrm_base = 0;
+ }
+ else {
+ SSize_t base; /* Temporary */
+
+ /* We have both: m * strlen(longer) + b = x_len_longer
+ * m * strlen(shorter) + b = x_len_shorter;
+ * subtracting yields:
+ * m * (strlen(longer) - strlen(shorter))
+ * = x_len_longer - x_len_shorter
+ * But we have set things up so that 'shorter' is 1 byte smaller
+ * than 'longer'. Hence:
+ * m = x_len_longer - x_len_shorter
+ *
+ * But if something went wrong, make sure the multiplier is at
+ * least 1.
+ */
+ if (x_len_longer > x_len_shorter) {
+ PL_collxfrm_mult = (STRLEN) x_len_longer - x_len_shorter;
+ }
+ else {
+ PL_collxfrm_mult = 1;
+ }
+
+ /* mx + b = len
+ * so: b = len - mx
+ * but in case something has gone wrong, make sure it is
+ * non-negative */
+ base = x_len_longer - PL_collxfrm_mult * (sizeof(longer) - 1);
+ if (base < 0) {
+ base = 0;
+ }
+
+ /* Add 1 for the trailing NUL */
+ PL_collxfrm_base = base + 1;
+ }
+
+#ifdef DEBUGGING
+ if (DEBUG_L_TEST || debug_initialization) {
+ PerlIO_printf(Perl_debug_log,
+ "%s:%d: ?UTF-8 locale=%d; x_len_shorter=%"UVuf", "
+ "x_len_longer=%"UVuf","
+ " collate multipler=%"UVuf", collate base=%"UVuf"\n",
+ __FILE__, __LINE__,
+ PL_in_utf8_COLLATE_locale,
+ x_len_shorter, x_len_longer,
+ PL_collxfrm_mult, PL_collxfrm_base);
+ }
+#endif
}
}
const char * const setlocale_init = (PerlEnv_getenv("PERL_SKIP_LOCALE_INIT"))
? NULL
: "";
-#ifdef DEBUGGING
- const bool debug = (PerlEnv_getenv("PERL_DEBUG_LOCALE_INIT"))
- ? TRUE
- : FALSE;
-# define DEBUG_LOCALE_INIT(category, locale, result) \
- STMT_START { \
- if (debug) { \
- PerlIO_printf(Perl_debug_log, \
- "%s:%d: %s\n", \
- __FILE__, __LINE__, \
- _setlocale_debug_string(category, \
- locale, \
- result)); \
- } \
- } STMT_END
-#else
-# define DEBUG_LOCALE_INIT(a,b,c)
-#endif
const char* trial_locales[5]; /* 5 = 1 each for "", LC_ALL, LANG, "", C */
unsigned int trial_locales_count;
const char * const lc_all = savepv(PerlEnv_getenv("LC_ALL"));
const char *system_default_locale = NULL;
#endif
+#ifdef DEBUGGING
+ debug_initialization = (PerlEnv_getenv("PERL_DEBUG_LOCALE_INIT"))
+ ? TRUE
+ : FALSE;
+# define DEBUG_LOCALE_INIT(category, locale, result) \
+ STMT_START { \
+ if (debug_initialization) { \
+ PerlIO_printf(Perl_debug_log, \
+ "%s:%d: %s\n", \
+ __FILE__, __LINE__, \
+ _setlocale_debug_string(category, \
+ locale, \
+ result)); \
+ } \
+ } STMT_END
+#else
+# define DEBUG_LOCALE_INIT(a,b,c)
+#endif
+
#ifndef LOCALE_ENVIRON_REQUIRED
PERL_UNUSED_VAR(done);
PERL_UNUSED_VAR(locale_param);
PERL_UNUSED_ARG(printwarn);
#endif /* USE_LOCALE */
+#ifdef DEBUGGING
+ /* So won't continue to output stuff */
+ debug_initialization = FALSE;
+#endif
+
return ok;
}
-
#ifdef USE_LOCALE_COLLATE
-/*
- * mem_collxfrm() is a bit like strxfrm() but with two important
- * differences. First, it handles embedded NULs. Second, it allocates
- * a bit more memory than needed for the transformed data itself.
- * The real transformed data begins at offset sizeof(collationix).
- * *xlen is set to the length of that, and doesn't include the collation index
- * size.
- * Please see sv_collxfrm() to see how this is used.
- */
-
char *
-Perl_mem_collxfrm(pTHX_ const char *input_string,
- STRLEN len,
- STRLEN *xlen
+Perl__mem_collxfrm(pTHX_ const char *input_string,
+ STRLEN len, /* Length of 'input_string' */
+ STRLEN *xlen, /* Set to length of returned string
+ (not including the collation index
+ prefix) */
+ bool utf8 /* Is the input in UTF-8? */
)
{
+
+ /* _mem_collxfrm() is a bit like strxfrm() but with two important
+ * differences. First, it handles embedded NULs. Second, it allocates a bit
+ * more memory than needed for the transformed data itself. The real
+ * transformed data begins at offset COLLXFRM_HDR_LEN. *xlen is set to
+ * the length of that, and doesn't include the collation index size.
+ * Please see sv_collxfrm() to see how this is used. */
+
+#define COLLXFRM_HDR_LEN sizeof(PL_collation_ix)
+
char * s = (char *) input_string;
STRLEN s_strlen = strlen(input_string);
- char *xbuf;
- STRLEN xAlloc, xout; /* xalloc is a reserved word in VC */
+ char *xbuf = NULL;
+ STRLEN xAlloc; /* xalloc is a reserved word in VC */
+ bool first_time = TRUE; /* Cleared after first loop iteration */
+
+ PERL_ARGS_ASSERT__MEM_COLLXFRM;
+
+ /* Must be NUL-terminated */
+ assert(*(input_string + len) == '\0');
- PERL_ARGS_ASSERT_MEM_COLLXFRM;
+ /* If this locale has defective collation, skip */
+ if (PL_collxfrm_base == 0 && PL_collxfrm_mult == 0) {
+ goto bad;
+ }
/* Replace any embedded NULs with the control that sorts before any others.
* This will give as good as possible results on strings that don't
* locale, find it */
if (*PL_strxfrm_min_char == '\0') {
int j;
- char * cur_min_x = NULL; /* Cur cp's xfrm, (except it also
+#ifdef DEBUGGING
+ U8 cur_min_cp = 1; /* The code point that sorts lowest, so far */
+#endif
+ char * cur_min_x = NULL; /* And its xfrm, (except it also
includes the collation index
prefixed. */
/* If needs to be 2 bytes, find them */
if (! UVCHR_IS_INVARIANT(j)) {
- continue; /* Can't handle variants yet */
+ char * d = cur_source;
+ append_utf8_from_native_byte((U8) j, (U8 **) &d);
+ trial_len = 2;
}
}
else if (! isCNTRL_LC(j)) {
}
/* Then transform it */
- x = mem_collxfrm(cur_source, trial_len, &x_len);
+ x = _mem_collxfrm(cur_source, trial_len, &x_len,
+ PL_in_utf8_COLLATE_locale);
/* If something went wrong (which it shouldn't), just
* ignore this code point */
if ( x_len == 0
- || strlen(x + sizeof(PL_collation_ix)) < x_len)
+ || strlen(x + COLLXFRM_HDR_LEN) < x_len)
{
continue;
}
/* If this character's transformation is lower than
* the current lowest, this one becomes the lowest */
if ( cur_min_x == NULL
- || strLT(x + sizeof(PL_collation_ix),
- cur_min_x + sizeof(PL_collation_ix)))
+ || strLT(x + COLLXFRM_HDR_LEN,
+ cur_min_x + COLLXFRM_HDR_LEN))
{
strcpy(PL_strxfrm_min_char, cur_source);
cur_min_x = x;
+#ifdef DEBUGGING
+ cur_min_cp = j;
+#endif
}
else {
Safefree(x);
* locale, arbitrarily use \001 */
if (cur_min_x == NULL) {
STRLEN x_len; /* temporary */
- cur_min_x = mem_collxfrm("\001", 1, &x_len);
+ cur_min_x = _mem_collxfrm("\001", 1, &x_len,
+ PL_in_utf8_COLLATE_locale);
/* cur_min_cp was already initialized to 1 */
}
+ DEBUG_L(PerlIO_printf(Perl_debug_log,
+ "_mem_collxfrm: lowest collating control in the 0-255 "
+ "range in locale %s is 0x%02X\n",
+ PL_collation_name,
+ cur_min_cp));
+ if (DEBUG_Lv_TEST) {
+ unsigned i;
+ PerlIO_printf(Perl_debug_log, "Its xfrm is");
+ for (i = 0; i < strlen(cur_min_x + COLLXFRM_HDR_LEN); i ++) {
+ PerlIO_printf(Perl_debug_log, " %02x",
+ (U8) *(cur_min_x + COLLXFRM_HDR_LEN + i));
+ }
+ PerlIO_printf(Perl_debug_log, "\n");
+ }
+
Safefree(cur_min_x);
}
len = strlen(s);
}
+ /* Make sure the UTF8ness of the string and locale match */
+ if (utf8 != PL_in_utf8_COLLATE_locale) {
+ const char * const t = s; /* Temporary so we can later find where the
+ input was */
+
+ /* Here they don't match. Change the string's to be what the locale is
+ * expecting */
+
+ if (! utf8) { /* locale is UTF-8, but input isn't; upgrade the input */
+ s = (char *) bytes_to_utf8((const U8 *) s, &len);
+ utf8 = TRUE;
+ }
+ else { /* locale is not UTF-8; but input is; downgrade the input */
+
+ s = (char *) bytes_from_utf8((const U8 *) s, &len, &utf8);
+
+ /* If the downgrade was successful we are done, but if the input
+ * contains things that require UTF-8 to represent, have to do
+ * damage control ... */
+ if (UNLIKELY(utf8)) {
+
+ /* What we do is construct a non-UTF-8 string with
+ * 1) the characters representable by a single byte converted
+ * to be so (if necessary);
+ * 2) and the rest converted to collate the same as the
+ * highest collating representable character. That makes
+ * them collate at the end. This is similar to how we
+ * handle embedded NULs, but we use the highest collating
+ * code point instead of the smallest. Like the NUL case,
+ * this isn't perfect, but is the best we can reasonably
+ * do. Every above-255 code point will sort the same as
+ * the highest-sorting 0-255 code point. If that code
+ * point can combine in a sequence with some other code
+ * points for weight calculations, us changing something to
+ * be it can adversely affect the results. But in most
+ * cases, it should work reasonably. And note that this is
+ * really an illegal situation: using code points above 255
+ * on a locale where only 0-255 are valid. If two strings
+ * sort entirely equal, then the sort order for the
+ * above-255 code points will be in code point order. */
+
+ utf8 = FALSE;
+
+ /* If we haven't calculated the code point with the maximum
+ * collating order for this locale, do so now */
+ if (! PL_strxfrm_max_cp) {
+ int j;
+
+ /* The current transformed string that collates the
+ * highest (except it also includes the prefixed collation
+ * index. */
+ char * cur_max_x = NULL;
+
+ /* Look through all legal code points (NUL isn't) */
+ for (j = 1; j < 256; j++) {
+ char * x;
+ STRLEN x_len;
+
+ /* Create a 1-char string of the current code point. */
+ char cur_source[] = { (char) j, '\0' };
+
+ /* Then transform it */
+ x = _mem_collxfrm(cur_source, 1, &x_len, FALSE);
+
+ /* If something went wrong (which it shouldn't), just
+ * ignore this code point */
+ if (x_len == 0) {
+ Safefree(x);
+ continue;
+ }
+
+ /* If this character's transformation is higher than
+ * the current highest, this one becomes the highest */
+ if ( cur_max_x == NULL
+ || strGT(x + COLLXFRM_HDR_LEN,
+ cur_max_x + COLLXFRM_HDR_LEN))
+ {
+ PL_strxfrm_max_cp = j;
+ cur_max_x = x;
+ }
+ else {
+ Safefree(x);
+ }
+ }
+
+ DEBUG_L(PerlIO_printf(Perl_debug_log,
+ "_mem_collxfrm: highest 1-byte collating character"
+ " in locale %s is 0x%02X\n",
+ PL_collation_name,
+ PL_strxfrm_max_cp));
+ if (DEBUG_Lv_TEST) {
+ unsigned i;
+ PerlIO_printf(Perl_debug_log, "Its xfrm is ");
+ for (i = 0;
+ i < strlen(cur_max_x + COLLXFRM_HDR_LEN);
+ i++)
+ {
+ PerlIO_printf(Perl_debug_log, " %02x",
+ (U8) cur_max_x[i + COLLXFRM_HDR_LEN]);
+ }
+ PerlIO_printf(Perl_debug_log, "\n");
+ }
+
+ Safefree(cur_max_x);
+ }
+
+ /* Here we know which legal code point collates the highest.
+ * We are ready to construct the non-UTF-8 string. The length
+ * will be at least 1 byte smaller than the input string
+ * (because we changed at least one 2-byte character into a
+ * single byte), but that is eaten up by the trailing NUL */
+ Newx(s, len, char);
+
+ {
+ STRLEN i;
+ STRLEN d= 0;
+
+ for (i = 0; i < len; i+= UTF8SKIP(t + i)) {
+ U8 cur_char = t[i];
+ if (UTF8_IS_INVARIANT(cur_char)) {
+ s[d++] = cur_char;
+ }
+ else if (UTF8_IS_DOWNGRADEABLE_START(cur_char)) {
+ s[d++] = EIGHT_BIT_UTF8_TO_NATIVE(cur_char, t[i+1]);
+ }
+ else { /* Replace illegal cp with highest collating
+ one */
+ s[d++] = PL_strxfrm_max_cp;
+ }
+ }
+ s[d++] = '\0';
+ Renew(s, d, char); /* Free up unused space */
+ }
+ }
+ }
+
+ /* Here, we have constructed a modified version of the input. It could
+ * be that we already had a modified copy before we did this version.
+ * If so, that copy is no longer needed */
+ if (t != input_string) {
+ Safefree(t);
+ }
+ }
+
/* The first element in the output is the collation id, used by
* sv_collxfrm(); then comes the space for the transformed string. The
* equation should give us a good estimate as to how much is needed */
- xAlloc = sizeof(PL_collation_ix) + PL_collxfrm_base + (PL_collxfrm_mult * len) + 1;
+ xAlloc = COLLXFRM_HDR_LEN
+ + PL_collxfrm_base
+ + (PL_collxfrm_mult * ((utf8)
+ ? utf8_length((U8 *) s, (U8 *) s + len)
+ : len));
Newx(xbuf, xAlloc, char);
if (UNLIKELY(! xbuf))
goto bad;
/* Store the collation id */
*(U32*)xbuf = PL_collation_ix;
- xout = sizeof(PL_collation_ix);
/* Then the transformation of the input. We loop until successful, or we
* give up */
for (;;) {
- STRLEN xused = strxfrm(xbuf + xout, s, xAlloc - xout);
+ *xlen = strxfrm(xbuf + COLLXFRM_HDR_LEN, s, xAlloc - COLLXFRM_HDR_LEN);
/* If the transformed string occupies less space than we told strxfrm()
* was available, it means it successfully transformed the whole
* string. */
- if (xused < xAlloc - xout) {
- xout += xused;
+ if (*xlen < xAlloc - COLLXFRM_HDR_LEN) {
break;
}
- if (UNLIKELY(xused >= PERL_INT_MAX))
+ if (UNLIKELY(*xlen >= PERL_INT_MAX))
goto bad;
- /* Otherwise it should be that the transformation stopped in the middle
- * because it ran out of space. Malloc more, and try again. */
- xAlloc = (2 * xAlloc) + 1;
+ /* A well-behaved strxfrm() returns exactly how much space it needs
+ * (not including the trailing NUL) when it fails due to not enough
+ * space being provided. Assume that this is the case unless it's been
+ * proven otherwise */
+ if (LIKELY(PL_strxfrm_is_behaved) && first_time) {
+ xAlloc = *xlen + COLLXFRM_HDR_LEN + 1;
+ }
+ else { /* Here, either:
+ * 1) The strxfrm() has previously shown bad behavior; or
+ * 2) It isn't the first time through the loop, which means
+ * that the strxfrm() is now showing bad behavior, because
+ * we gave it what it said was needed in the previous
+ * iteration, and it came back saying it needed still more.
+ * (Many versions of cygwin fit this. When the buffer size
+ * isn't sufficient, they return the input size instead of
+ * how much is needed.)
+ * Increase the buffer size by a fixed percentage and try again. */
+ xAlloc += (xAlloc / 4) + 1;
+ PL_strxfrm_is_behaved = FALSE;
+
+#ifdef DEBUGGING
+ if (DEBUG_Lv_TEST || debug_initialization) {
+ PerlIO_printf(Perl_debug_log,
+ "_mem_collxfrm required more space than previously calculated"
+ " for locale %s, trying again with new guess=%d+%"UVuf"\n",
+ PL_collation_name, (int) COLLXFRM_HDR_LEN,
+ (UV) xAlloc - COLLXFRM_HDR_LEN);
+ }
+#endif
+ }
+
Renew(xbuf, xAlloc, char);
if (UNLIKELY(! xbuf))
goto bad;
+
+ first_time = FALSE;
}
- *xlen = xout - sizeof(PL_collation_ix);
+#ifdef DEBUGGING
+ if (DEBUG_Lv_TEST || debug_initialization) {
+ unsigned i;
+ PerlIO_printf(Perl_debug_log,
+ "_mem_collxfrm[%d]: returning %"UVuf" for locale %s '%s'\n",
+ PL_collation_ix, *xlen, PL_collation_name, input_string);
+ PerlIO_printf(Perl_debug_log, "Its xfrm is");
+ for (i = COLLXFRM_HDR_LEN; i < *xlen + COLLXFRM_HDR_LEN; i++) {
+ PerlIO_printf(Perl_debug_log, " %02x", (U8) xbuf[i]);
+ }
+ PerlIO_printf(Perl_debug_log, "\n");
+ }
+#endif
+
+ /* Free up unneeded space; retain ehough for trailing NUL */
+ Renew(xbuf, COLLXFRM_HDR_LEN + *xlen + 1, char);
if (s != input_string) {
Safefree(s);
Safefree(s);
}
*xlen = 0;
+#ifdef DEBUGGING
+ if (DEBUG_Lv_TEST || debug_initialization) {
+ PerlIO_printf(Perl_debug_log, "_mem_collxfrm[%d] returning NULL\n",
+ PL_collation_ix);
+ }
+#endif
return NULL;
}