2020-08-11 17:10:23 +08:00
|
|
|
// © 2016 and later: Unicode, Inc. and others.
|
|
|
|
// License & terms of use: http://www.unicode.org/copyright.html
|
|
|
|
/*
|
|
|
|
**********************************************************************
|
|
|
|
* Copyright (c) 2001-2011, International Business Machines
|
|
|
|
* Corporation and others. All Rights Reserved.
|
|
|
|
**********************************************************************
|
|
|
|
* Date Name Description
|
|
|
|
* 11/19/2001 aliu Creation.
|
|
|
|
**********************************************************************
|
|
|
|
*/
|
|
|
|
|
|
|
|
#ifndef ICU_UTIL_H
|
|
|
|
#define ICU_UTIL_H
|
|
|
|
|
2021-04-22 20:08:59 +08:00
|
|
|
#include "charstr.h"
|
2020-08-11 17:10:23 +08:00
|
|
|
#include "unicode/unistr.h"
|
2021-04-22 20:08:59 +08:00
|
|
|
#include "unicode/uobject.h"
|
|
|
|
#include "unicode/utypes.h"
|
2020-08-11 17:10:23 +08:00
|
|
|
//--------------------------------------------------------------------
|
|
|
|
// class ICU_Utility
|
|
|
|
// i18n utility functions, scoped into the class ICU_Utility.
|
|
|
|
//--------------------------------------------------------------------
|
|
|
|
|
|
|
|
U_NAMESPACE_BEGIN
|
|
|
|
|
|
|
|
class UnicodeMatcher;
|
|
|
|
|
|
|
|
class U_COMMON_API ICU_Utility /* not : public UObject because all methods are static */ {
|
|
|
|
public:
|
|
|
|
|
|
|
|
/**
|
|
|
|
* Append a number to the given UnicodeString in the given radix.
|
|
|
|
* Standard digits '0'-'9' are used and letters 'A'-'Z' for
|
|
|
|
* radices 11 through 36.
|
|
|
|
* @param result the digits of the number are appended here
|
|
|
|
* @param n the number to be converted to digits; may be negative.
|
|
|
|
* If negative, a '-' is prepended to the digits.
|
|
|
|
* @param radix a radix from 2 to 36 inclusive.
|
|
|
|
* @param minDigits the minimum number of digits, not including
|
|
|
|
* any '-', to produce. Values less than 2 have no effect. One
|
|
|
|
* digit is always emitted regardless of this parameter.
|
|
|
|
* @return a reference to result
|
|
|
|
*/
|
|
|
|
static UnicodeString& appendNumber(UnicodeString& result, int32_t n,
|
|
|
|
int32_t radix = 10,
|
|
|
|
int32_t minDigits = 1);
|
|
|
|
|
|
|
|
/** Returns a bogus UnicodeString by value. */
|
|
|
|
static inline UnicodeString makeBogusString() {
|
|
|
|
UnicodeString result;
|
|
|
|
result.setToBogus();
|
|
|
|
return result;
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* Return true if the character is NOT printable ASCII.
|
2021-10-28 14:15:28 +08:00
|
|
|
* The tab, newline and linefeed characters are considered unprintable.
|
2020-08-11 17:10:23 +08:00
|
|
|
*/
|
|
|
|
static UBool isUnprintable(UChar32 c);
|
|
|
|
|
|
|
|
/**
|
2021-10-28 14:15:28 +08:00
|
|
|
* @return true for control codes and for surrogate and noncharacter code points
|
|
|
|
*/
|
|
|
|
static UBool shouldAlwaysBeEscaped(UChar32 c);
|
|
|
|
|
|
|
|
/**
|
|
|
|
* Escapes one unprintable code point using \uxxxx notation for U+0000 to
|
2020-08-11 17:10:23 +08:00
|
|
|
* U+FFFF and \Uxxxxxxxx for U+10000 and above. If the character is
|
|
|
|
* printable ASCII, then do nothing and return false. Otherwise,
|
|
|
|
* append the escaped notation and return true.
|
|
|
|
*/
|
|
|
|
static UBool escapeUnprintable(UnicodeString& result, UChar32 c);
|
|
|
|
|
2021-10-28 14:15:28 +08:00
|
|
|
/**
|
|
|
|
* Escapes one code point using \uxxxx notation
|
|
|
|
* for U+0000 to U+FFFF and \Uxxxxxxxx for U+10000 and above.
|
|
|
|
* @return result
|
|
|
|
*/
|
|
|
|
static UnicodeString &escape(UnicodeString& result, UChar32 c);
|
|
|
|
|
2020-08-11 17:10:23 +08:00
|
|
|
/**
|
|
|
|
* Returns the index of a character, ignoring quoted text.
|
|
|
|
* For example, in the string "abc'hide'h", the 'h' in "hide" will not be
|
|
|
|
* found by a search for 'h'.
|
|
|
|
* @param text text to be searched
|
|
|
|
* @param start the beginning index, inclusive; <code>0 <= start
|
|
|
|
* <= limit</code>.
|
|
|
|
* @param limit the ending index, exclusive; <code>start <= limit
|
|
|
|
* <= text.length()</code>.
|
|
|
|
* @param c character to search for
|
|
|
|
* @return Offset of the first instance of c, or -1 if not found.
|
|
|
|
*/
|
|
|
|
//?FOR FUTURE USE. DISABLE FOR NOW for coverage reasons.
|
|
|
|
// static int32_t quotedIndexOf(const UnicodeString& text,
|
|
|
|
// int32_t start, int32_t limit,
|
2023-05-23 08:05:01 +08:00
|
|
|
// char16_t c);
|
2020-08-11 17:10:23 +08:00
|
|
|
|
|
|
|
/**
|
|
|
|
* Skip over a sequence of zero or more white space characters at pos.
|
|
|
|
* @param advance if true, advance pos to the first non-white-space
|
|
|
|
* character at or after pos, or str.length(), if there is none.
|
|
|
|
* Otherwise leave pos unchanged.
|
|
|
|
* @return the index of the first non-white-space character at or
|
|
|
|
* after pos, or str.length(), if there is none.
|
|
|
|
*/
|
|
|
|
static int32_t skipWhitespace(const UnicodeString& str, int32_t& pos,
|
|
|
|
UBool advance = false);
|
|
|
|
|
|
|
|
/**
|
|
|
|
* Skip over Pattern_White_Space in a Replaceable.
|
|
|
|
* Skipping may be done in the forward or
|
|
|
|
* reverse direction. In either case, the leftmost index will be
|
|
|
|
* inclusive, and the rightmost index will be exclusive. That is,
|
|
|
|
* given a range defined as [start, limit), the call
|
|
|
|
* skipWhitespace(text, start, limit) will advance start past leading
|
|
|
|
* whitespace, whereas the call skipWhitespace(text, limit, start),
|
|
|
|
* will back up limit past trailing whitespace.
|
|
|
|
* @param text the text to be analyzed
|
|
|
|
* @param pos either the start or limit of a range of 'text', to skip
|
|
|
|
* leading or trailing whitespace, respectively
|
|
|
|
* @param stop either the limit or start of a range of 'text', to skip
|
|
|
|
* leading or trailing whitespace, respectively
|
|
|
|
* @return the new start or limit, depending on what was passed in to
|
|
|
|
* 'pos'
|
|
|
|
*/
|
|
|
|
//?FOR FUTURE USE. DISABLE FOR NOW for coverage reasons.
|
|
|
|
//? static int32_t skipWhitespace(const Replaceable& text,
|
|
|
|
//? int32_t pos, int32_t stop);
|
|
|
|
|
|
|
|
/**
|
|
|
|
* Parse a single non-whitespace character 'ch', optionally
|
|
|
|
* preceded by whitespace.
|
|
|
|
* @param id the string to be parsed
|
|
|
|
* @param pos INPUT-OUTPUT parameter. On input, pos[0] is the
|
|
|
|
* offset of the first character to be parsed. On output, pos[0]
|
|
|
|
* is the index after the last parsed character. If the parse
|
|
|
|
* fails, pos[0] will be unchanged.
|
|
|
|
* @param ch the non-whitespace character to be parsed.
|
|
|
|
* @return true if 'ch' is seen preceded by zero or more
|
|
|
|
* whitespace characters.
|
|
|
|
*/
|
2023-05-23 08:05:01 +08:00
|
|
|
static UBool parseChar(const UnicodeString& id, int32_t& pos, char16_t ch);
|
2020-08-11 17:10:23 +08:00
|
|
|
|
|
|
|
/**
|
|
|
|
* Parse a pattern string starting at offset pos. Keywords are
|
|
|
|
* matched case-insensitively. Spaces may be skipped and may be
|
|
|
|
* optional or required. Integer values may be parsed, and if
|
|
|
|
* they are, they will be returned in the given array. If
|
|
|
|
* successful, the offset of the next non-space character is
|
|
|
|
* returned. On failure, -1 is returned.
|
|
|
|
* @param pattern must only contain lowercase characters, which
|
|
|
|
* will match their uppercase equivalents as well. A space
|
|
|
|
* character matches one or more required spaces. A '~' character
|
|
|
|
* matches zero or more optional spaces. A '#' character matches
|
|
|
|
* an integer and stores it in parsedInts, which the caller must
|
|
|
|
* ensure has enough capacity.
|
|
|
|
* @param parsedInts array to receive parsed integers. Caller
|
|
|
|
* must ensure that parsedInts.length is >= the number of '#'
|
|
|
|
* signs in 'pattern'.
|
|
|
|
* @return the position after the last character parsed, or -1 if
|
|
|
|
* the parse failed
|
|
|
|
*/
|
|
|
|
static int32_t parsePattern(const UnicodeString& rule, int32_t pos, int32_t limit,
|
|
|
|
const UnicodeString& pattern, int32_t* parsedInts);
|
|
|
|
|
|
|
|
/**
|
|
|
|
* Parse a pattern string within the given Replaceable and a parsing
|
|
|
|
* pattern. Characters are matched literally and case-sensitively
|
|
|
|
* except for the following special characters:
|
|
|
|
*
|
|
|
|
* ~ zero or more Pattern_White_Space chars
|
|
|
|
*
|
|
|
|
* If end of pattern is reached with all matches along the way,
|
|
|
|
* pos is advanced to the first unparsed index and returned.
|
|
|
|
* Otherwise -1 is returned.
|
|
|
|
* @param pat pattern that controls parsing
|
|
|
|
* @param text text to be parsed, starting at index
|
|
|
|
* @param index offset to first character to parse
|
|
|
|
* @param limit offset after last character to parse
|
|
|
|
* @return index after last parsed character, or -1 on parse failure.
|
|
|
|
*/
|
|
|
|
static int32_t parsePattern(const UnicodeString& pat,
|
|
|
|
const Replaceable& text,
|
|
|
|
int32_t index,
|
|
|
|
int32_t limit);
|
|
|
|
|
|
|
|
/**
|
|
|
|
* Parse an integer at pos, either of the form \d+ or of the form
|
|
|
|
* 0x[0-9A-Fa-f]+ or 0[0-7]+, that is, in standard decimal, hex,
|
|
|
|
* or octal format.
|
|
|
|
* @param pos INPUT-OUTPUT parameter. On input, the index of the first
|
|
|
|
* character to parse. On output, the index of the character after the
|
|
|
|
* last parsed character.
|
|
|
|
*/
|
|
|
|
static int32_t parseInteger(const UnicodeString& rule, int32_t& pos, int32_t limit);
|
|
|
|
|
|
|
|
/**
|
|
|
|
* Parse an integer at pos using only ASCII digits.
|
|
|
|
* Base 10 only.
|
|
|
|
* @param pos INPUT-OUTPUT parameter. On input, the index of the first
|
|
|
|
* character to parse. On output, the index of the character after the
|
|
|
|
* last parsed character.
|
|
|
|
*/
|
|
|
|
static int32_t parseAsciiInteger(const UnicodeString& str, int32_t& pos);
|
|
|
|
|
|
|
|
/**
|
|
|
|
* Parse a Unicode identifier from the given string at the given
|
|
|
|
* position. Return the identifier, or an empty string if there
|
|
|
|
* is no identifier.
|
|
|
|
* @param str the string to parse
|
2021-10-28 14:15:28 +08:00
|
|
|
* @param pos INPUT-OUTPUT parameter. On INPUT, pos is the
|
2020-08-11 17:10:23 +08:00
|
|
|
* first character to examine. It must be less than str.length(),
|
|
|
|
* and it must not point to a whitespace character. That is, must
|
|
|
|
* have pos < str.length() and
|
|
|
|
* !UCharacter::isWhitespace(str.char32At(pos)). On
|
|
|
|
* OUTPUT, the position after the last parsed character.
|
|
|
|
* @return the Unicode identifier, or an empty string if there is
|
|
|
|
* no valid identifier at pos.
|
|
|
|
*/
|
|
|
|
static UnicodeString parseUnicodeIdentifier(const UnicodeString& str, int32_t& pos);
|
|
|
|
|
|
|
|
/**
|
|
|
|
* Parse an unsigned 31-bit integer at the given offset. Use
|
|
|
|
* UCharacter.digit() to parse individual characters into digits.
|
|
|
|
* @param text the text to be parsed
|
|
|
|
* @param pos INPUT-OUTPUT parameter. On entry, pos is the
|
|
|
|
* offset within text at which to start parsing; it should point
|
|
|
|
* to a valid digit. On exit, pos is the offset after the last
|
|
|
|
* parsed character. If the parse failed, it will be unchanged on
|
|
|
|
* exit. Must be >= 0 on entry.
|
|
|
|
* @param radix the radix in which to parse; must be >= 2 and <=
|
|
|
|
* 36.
|
|
|
|
* @return a non-negative parsed number, or -1 upon parse failure.
|
|
|
|
* Parse fails if there are no digits, that is, if pos does not
|
|
|
|
* point to a valid digit on entry, or if the number to be parsed
|
|
|
|
* does not fit into a 31-bit unsigned integer.
|
|
|
|
*/
|
|
|
|
static int32_t parseNumber(const UnicodeString& text,
|
|
|
|
int32_t& pos, int8_t radix);
|
|
|
|
|
|
|
|
static void appendToRule(UnicodeString& rule,
|
|
|
|
UChar32 c,
|
|
|
|
UBool isLiteral,
|
|
|
|
UBool escapeUnprintable,
|
|
|
|
UnicodeString& quoteBuf);
|
|
|
|
|
|
|
|
static void appendToRule(UnicodeString& rule,
|
|
|
|
const UnicodeString& text,
|
|
|
|
UBool isLiteral,
|
|
|
|
UBool escapeUnprintable,
|
|
|
|
UnicodeString& quoteBuf);
|
|
|
|
|
|
|
|
static void appendToRule(UnicodeString& rule,
|
|
|
|
const UnicodeMatcher* matcher,
|
|
|
|
UBool escapeUnprintable,
|
|
|
|
UnicodeString& quoteBuf);
|
|
|
|
|
|
|
|
private:
|
|
|
|
// do not instantiate
|
2022-10-28 14:11:55 +08:00
|
|
|
ICU_Utility() = delete;
|
2020-08-11 17:10:23 +08:00
|
|
|
};
|
|
|
|
|
|
|
|
U_NAMESPACE_END
|
|
|
|
|
|
|
|
#endif
|
|
|
|
//eof
|