ICU-21480 Update double-conversion

Author: sffc

icu4c/source/i18n/double-conversion-double-to-string.cpp

@@ -107,19 +107,19 @@ void DoubleToStringConverter::CreateExponentialRepresentation(
       result_builder->AddCharacter('+');
     }
   }
-  if (exponent == 0) {
-    result_builder->AddCharacter('0');
-    return;
-  }
   DOUBLE_CONVERSION_ASSERT(exponent < 1e4);
   // Changing this constant requires updating the comment of DoubleToStringConverter constructor
   const int kMaxExponentLength = 5;
   char buffer[kMaxExponentLength + 1];
   buffer[kMaxExponentLength] = '\0';
   int first_char_pos = kMaxExponentLength;
-  while (exponent > 0) {
-    buffer[--first_char_pos] = '0' + (exponent % 10);
-    exponent /= 10;
+  if (exponent == 0) {
+    buffer[--first_char_pos] = '0';
+  } else {
+    while (exponent > 0) {
+      buffer[--first_char_pos] = '0' + (exponent % 10);
+      exponent /= 10;
+    }
   }
   // Add prefix '0' to make exponent width >= min(min_exponent_with_, kMaxExponentLength)
   // For example: convert 1e+9 -> 1e+09, if min_exponent_with_ is set to 2
@@ -342,9 +342,21 @@ bool DoubleToStringConverter::ToPrecision(double value,
   int exponent = decimal_point - 1;
 
   int extra_zero = ((flags_ & EMIT_TRAILING_ZERO_AFTER_POINT) != 0) ? 1 : 0;
-  if ((-decimal_point + 1 > max_leading_padding_zeroes_in_precision_mode_) ||
+  bool as_exponential =
+      (-decimal_point + 1 > max_leading_padding_zeroes_in_precision_mode_) ||
       (decimal_point - precision + extra_zero >
-       max_trailing_padding_zeroes_in_precision_mode_)) {
+       max_trailing_padding_zeroes_in_precision_mode_);
+  if ((flags_ & NO_TRAILING_ZERO) != 0) {
+    // Truncate trailing zeros that occur after the decimal point (if exponential,
+    // that is everything after the first digit).
+    int stop = as_exponential ? 1 : std::max(1, decimal_point);
+    while (decimal_rep_length > stop && decimal_rep[decimal_rep_length - 1] == '0') {
+      --decimal_rep_length;
+    }
+    // Clamp precision to avoid the code below re-adding the zeros.
+    precision = std::min(precision, decimal_rep_length);
+  }
+  if (as_exponential) {
     // Fill buffer to contain 'precision' digits.
     // Usually the buffer is already at the correct length, but 'DoubleToAscii'
     // is allowed to return less characters.

icu4c/source/i18n/double-conversion-double-to-string.h

@@ -48,12 +48,11 @@ namespace double_conversion {
 
 class DoubleToStringConverter {
  public:
-#if 0 // not needed for ICU
   // When calling ToFixed with a double > 10^kMaxFixedDigitsBeforePoint
   // or a requested_digits parameter > kMaxFixedDigitsAfterPoint then the
   // function returns false.
   static const int kMaxFixedDigitsBeforePoint = 60;
-  static const int kMaxFixedDigitsAfterPoint = 60;
+  static const int kMaxFixedDigitsAfterPoint = 100;
 
   // When calling ToExponential with a requested_digits
   // parameter > kMaxExponentialDigits then the function returns false.
@@ -65,12 +64,36 @@ class DoubleToStringConverter {
   static const int kMinPrecisionDigits = 1;
   static const int kMaxPrecisionDigits = 120;
 
+  // The maximal number of digits that are needed to emit a double in base 10.
+  // A higher precision can be achieved by using more digits, but the shortest
+  // accurate representation of any double will never use more digits than
+  // kBase10MaximalLength.
+  // Note that DoubleToAscii null-terminates its input. So the given buffer
+  // should be at least kBase10MaximalLength + 1 characters long.
+  static const int kBase10MaximalLength = 17;
+
+  // The maximal number of digits that are needed to emit a single in base 10.
+  // A higher precision can be achieved by using more digits, but the shortest
+  // accurate representation of any single will never use more digits than
+  // kBase10MaximalLengthSingle.
+  static const int kBase10MaximalLengthSingle = 9;
+
+  // The length of the longest string that 'ToShortest' can produce when the
+  // converter is instantiated with EcmaScript defaults (see
+  // 'EcmaScriptConverter')
+  // This value does not include the trailing '\0' character.
+  // This amount of characters is needed for negative values that hit the
+  // 'decimal_in_shortest_low' limit. For example: "-0.0000033333333333333333"
+  static const int kMaxCharsEcmaScriptShortest = 25;
+
+#if 0 // not needed for ICU
   enum Flags {
     NO_FLAGS = 0,
     EMIT_POSITIVE_EXPONENT_SIGN = 1,
     EMIT_TRAILING_DECIMAL_POINT = 2,
     EMIT_TRAILING_ZERO_AFTER_POINT = 4,
-    UNIQUE_ZERO = 8
+    UNIQUE_ZERO = 8,
+    NO_TRAILING_ZERO = 16
   };
 
   // Flags should be a bit-or combination of the possible Flags-enum.
@@ -82,9 +105,13 @@ class DoubleToStringConverter {
   //    Example: 2345.0 is converted to "2345.".
   //  - EMIT_TRAILING_ZERO_AFTER_POINT: in addition to a trailing decimal point
   //    emits a trailing '0'-character. This flag requires the
-  //    EXMIT_TRAILING_DECIMAL_POINT flag.
+  //    EMIT_TRAILING_DECIMAL_POINT flag.
   //    Example: 2345.0 is converted to "2345.0".
   //  - UNIQUE_ZERO: "-0.0" is converted to "0.0".
+  //  - NO_TRAILING_ZERO: Trailing zeros are removed from the fractional portion
+  //    of the result in precision mode. Matches printf's %g.
+  //    When EMIT_TRAILING_ZERO_AFTER_POINT is also given, one trailing zero is
+  //    preserved.
   //
   // Infinity symbol and nan_symbol provide the string representation for these
   // special values. If the string is NULL and the special value is encountered
@@ -152,6 +179,14 @@ class DoubleToStringConverter {
   }
 
   // Returns a converter following the EcmaScript specification.
+  //
+  // Flags: UNIQUE_ZERO and EMIT_POSITIVE_EXPONENT_SIGN.
+  // Special values: "Infinity" and "NaN".
+  // Lower case 'e' for exponential values.
+  // decimal_in_shortest_low: -6
+  // decimal_in_shortest_high: 21
+  // max_leading_padding_zeroes_in_precision_mode: 6
+  // max_trailing_padding_zeroes_in_precision_mode: 0
   static const DoubleToStringConverter& EcmaScriptConverter();
 
   // Computes the shortest string of digits that correctly represent the input
@@ -177,6 +212,21 @@ class DoubleToStringConverter {
   // Returns true if the conversion succeeds. The conversion always succeeds
   // except when the input value is special and no infinity_symbol or
   // nan_symbol has been given to the constructor.
+  //
+  // The length of the longest result is the maximum of the length of the
+  // following string representations (each with possible examples):
+  // - NaN and negative infinity: "NaN", "-Infinity", "-inf".
+  // - -10^(decimal_in_shortest_high - 1):
+  //      "-100000000000000000000", "-1000000000000000.0"
+  // - the longest string in range [0; -10^decimal_in_shortest_low]. Generally,
+  //   this string is 3 + kBase10MaximalLength - decimal_in_shortest_low.
+  //   (Sign, '0', decimal point, padding zeroes for decimal_in_shortest_low,
+  //   and the significant digits).
+  //      "-0.0000033333333333333333", "-0.0012345678901234567"
+  // - the longest exponential representation. (A negative number with
+  //   kBase10MaximalLength significant digits).
+  //      "-1.7976931348623157e+308", "-1.7976931348623157E308"
+  // In addition, the buffer must be able to hold the trailing '\0' character.
   bool ToShortest(double value, StringBuilder* result_builder) const {
     return ToShortestIeeeNumber(value, result_builder, SHORTEST);
   }
@@ -217,9 +267,11 @@ class DoubleToStringConverter {
   //     been provided to the constructor,
   //   - 'value' > 10^kMaxFixedDigitsBeforePoint, or
   //   - 'requested_digits' > kMaxFixedDigitsAfterPoint.
-  // The last two conditions imply that the result will never contain more than
-  // 1 + kMaxFixedDigitsBeforePoint + 1 + kMaxFixedDigitsAfterPoint characters
+  // The last two conditions imply that the result for non-special values never
+  // contains more than
+  //  1 + kMaxFixedDigitsBeforePoint + 1 + kMaxFixedDigitsAfterPoint characters
   // (one additional character for the sign, and one for the decimal point).
+  // In addition, the buffer must be able to hold the trailing '\0' character.
   bool ToFixed(double value,
                int requested_digits,
                StringBuilder* result_builder) const;
@@ -248,14 +300,17 @@ class DoubleToStringConverter {
   //   - the input value is special and no infinity_symbol or nan_symbol has
   //     been provided to the constructor,
   //   - 'requested_digits' > kMaxExponentialDigits.
-  // The last condition implies that the result will never contain more than
+  //
+  // The last condition implies that the result never contains more than
   // kMaxExponentialDigits + 8 characters (the sign, the digit before the
   // decimal point, the decimal point, the exponent character, the
   // exponent's sign, and at most 3 exponent digits).
+  // In addition, the buffer must be able to hold the trailing '\0' character.
   bool ToExponential(double value,
                      int requested_digits,
                      StringBuilder* result_builder) const;
 
+
   // Computes 'precision' leading digits of the given 'value' and returns them
   // either in exponential or decimal format, depending on
   // max_{leading|trailing}_padding_zeroes_in_precision_mode (given to the
@@ -287,9 +342,11 @@ class DoubleToStringConverter {
   //     been provided to the constructor,
   //   - precision < kMinPericisionDigits
   //   - precision > kMaxPrecisionDigits
-  // The last condition implies that the result will never contain more than
+  //
+  // The last condition implies that the result never contains more than
   // kMaxPrecisionDigits + 7 characters (the sign, the decimal point, the
   // exponent character, the exponent's sign, and at most 3 exponent digits).
+  // In addition, the buffer must be able to hold the trailing '\0' character.
   bool ToPrecision(double value,
                    int precision,
                    StringBuilder* result_builder) const;
@@ -310,14 +367,6 @@ class DoubleToStringConverter {
     PRECISION
   };
 
-  // The maximal number of digits that are needed to emit a double in base 10.
-  // A higher precision can be achieved by using more digits, but the shortest
-  // accurate representation of any double will never use more digits than
-  // kBase10MaximalLength.
-  // Note that DoubleToAscii null-terminates its input. So the given buffer
-  // should be at least kBase10MaximalLength + 1 characters long.
-  static const int kBase10MaximalLength = 17;
-
   // Converts the given double 'v' to digit characters. 'v' must not be NaN,
   // +Infinity, or -Infinity. In SHORTEST_SINGLE-mode this restriction also
   // applies to 'v' after it has been casted to a single-precision float. That

icu4c/source/i18n/double-conversion-string-to-double.cpp

@@ -51,6 +51,18 @@
 // ICU PATCH: Wrap in ICU namespace
 U_NAMESPACE_BEGIN
 
+#ifdef _MSC_VER
+#  if _MSC_VER >= 1900
+// Fix MSVC >= 2015 (_MSC_VER == 1900) warning
+// C4244: 'argument': conversion from 'const uc16' to 'char', possible loss of data
+// against Advance and friends, when instantiated with **it as char, not uc16.
+ __pragma(warning(disable: 4244))
+#  endif
+#  if _MSC_VER <= 1700 // VS2012, see IsDecimalDigitForRadix warning fix, below
+#    define VS2012_RADIXWARN
+#  endif
+#endif
+
 namespace double_conversion {
 
 namespace {
@@ -170,9 +182,9 @@ static double SignedZero(bool sign) {
 //
 // The function is small and could be inlined, but VS2012 emitted a warning
 // because it constant-propagated the radix and concluded that the last
-// condition was always true. By moving it into a separate function the
-// compiler wouldn't warn anymore.
-#ifdef _MSC_VER
+// condition was always true. Moving it into a separate function and
+// suppressing optimisation keeps the compiler from warning.
+#ifdef VS2012_RADIXWARN
 #pragma optimize("",off)
 static bool IsDecimalDigitForRadix(int c, int radix) {
   return '0' <= c && c <= '9' && (c - '0') < radix;
@@ -738,11 +750,17 @@ double StringToDoubleConverter::StringToIeee(
   DOUBLE_CONVERSION_ASSERT(buffer_pos < kBufferSize);
   buffer[buffer_pos] = '\0';
 
+  // Code above ensures there are no leading zeros and the buffer has fewer than
+  // kMaxSignificantDecimalDigits characters. Trim trailing zeros.
+  Vector<const char> chars(buffer, buffer_pos);
+  chars = TrimTrailingZeros(chars);
+  exponent += buffer_pos - chars.length();
+
   double converted;
   if (read_as_double) {
-    converted = Strtod(Vector<const char>(buffer, buffer_pos), exponent);
+    converted = StrtodTrimmed(chars, exponent);
   } else {
-    converted = Strtof(Vector<const char>(buffer, buffer_pos), exponent);
+    converted = StrtofTrimmed(chars, exponent);
   }
   *processed_characters_count = static_cast<int>(current - input);
   return sign? -converted: converted;

icu4c/source/i18n/double-conversion-strtod.cpp

@@ -115,17 +115,6 @@ static Vector<const char> TrimLeadingZeros(Vector<const char> buffer) {
   return Vector<const char>(buffer.start(), 0);
 }
 
-
-static Vector<const char> TrimTrailingZeros(Vector<const char> buffer) {
-  for (int i = buffer.length() - 1; i >= 0; --i) {
-    if (buffer[i] != '0') {
-      return buffer.SubVector(0, i + 1);
-    }
-  }
-  return Vector<const char>(buffer.start(), 0);
-}
-
-
 static void CutToMaxSignificantDigits(Vector<const char> buffer,
                                        int exponent,
                                        char* significant_buffer,
@@ -216,12 +205,14 @@ static bool DoubleStrtod(Vector<const char> trimmed,
                          int exponent,
                          double* result) {
 #if !defined(DOUBLE_CONVERSION_CORRECT_DOUBLE_OPERATIONS)
+  // Avoid "unused parameter" warnings
+  (void) trimmed;
+  (void) exponent;
+  (void) result;
   // On x86 the floating-point stack can be 64 or 80 bits wide. If it is
   // 80 bits wide (as is the case on Linux) then double-rounding occurs and the
   // result is not accurate.
   // We know that Windows32 uses 64 bits and is therefore accurate.
-  // Note that the ARM simulator is compiled for 32bits. It therefore exhibits
-  // the same problem.
   return false;
 #else
   if (trimmed.length() <= kMaxExactDoubleIntegerDecimalDigits) {
@@ -473,6 +464,11 @@ static bool IsNonZeroDigit(const char d) {
   return ('1' <= d) && (d <= '9');
 }
 
+#ifdef __has_cpp_attribute
+#if __has_cpp_attribute(maybe_unused)
+[[maybe_unused]]
+#endif
+#endif
 static bool AssertTrimmedDigits(const Vector<const char>& buffer) {
   for(int i = 0; i < buffer.length(); ++i) {
     if(!IsDigit(buffer[i])) {
@@ -545,6 +541,12 @@ float Strtof(Vector<const char> buffer, int exponent) {
   TrimAndCut(buffer, exponent, copy_buffer, kMaxSignificantDecimalDigits,
              &trimmed, &updated_exponent);
   exponent = updated_exponent;
+  return StrtofTrimmed(trimmed, exponent);
+}
+
+float StrtofTrimmed(Vector<const char> trimmed, int exponent) {
+  DOUBLE_CONVERSION_ASSERT(trimmed.length() <= kMaxSignificantDecimalDigits);
+  DOUBLE_CONVERSION_ASSERT(AssertTrimmedDigits(trimmed));
 
   double double_guess;
   bool is_correct = ComputeGuess(trimmed, exponent, &double_guess);

icu4c/source/i18n/double-conversion-strtod.h

@@ -54,11 +54,25 @@ double Strtod(Vector<const char> buffer, int exponent);
 // contain a dot or a sign. It must not start with '0', and must not be empty.
 float Strtof(Vector<const char> buffer, int exponent);
 
-// For special use cases, the heart of the Strtod() function is also available
-// separately, it assumes that 'trimmed' is as produced by TrimAndCut(), i.e.
-// no leading or trailing zeros, also no lone zero, and not 'too many' digits.
+// Same as Strtod, but assumes that 'trimmed' is already trimmed, as if run
+// through TrimAndCut. That is, 'trimmed' must have no leading or trailing
+// zeros, must not be a lone zero, and must not have 'too many' digits.
 double StrtodTrimmed(Vector<const char> trimmed, int exponent);
 
+// Same as Strtof, but assumes that 'trimmed' is already trimmed, as if run
+// through TrimAndCut. That is, 'trimmed' must have no leading or trailing
+// zeros, must not be a lone zero, and must not have 'too many' digits.
+float StrtofTrimmed(Vector<const char> trimmed, int exponent);
+
+inline Vector<const char> TrimTrailingZeros(Vector<const char> buffer) {
+  for (int i = buffer.length() - 1; i >= 0; --i) {
+    if (buffer[i] != '0') {
+      return buffer.SubVector(0, i + 1);
+    }
+  }
+  return Vector<const char>(buffer.start(), 0);
+}
+
 }  // namespace double_conversion
 
 // ICU PATCH: Close ICU namespace

icu4c/source/i18n/double-conversion-utils.h

@@ -118,7 +118,7 @@ int main(int argc, char** argv) {
     defined(__ARMEL__) || defined(__avr32__) || defined(_M_ARM) || defined(_M_ARM64) || \
     defined(__hppa__) || defined(__ia64__) || \
     defined(__mips__) || \
-    defined(__nios2__) || \
+    defined(__nios2__) || defined(__ghs) || \
     defined(__powerpc__) || defined(__ppc__) || defined(__ppc64__) || \
     defined(_POWER) || defined(_ARCH_PPC) || defined(_ARCH_PPC64) || \
     defined(__sparc__) || defined(__sparc) || defined(__s390__) || \

vendor/double-conversion/upstream/.gitignore

@@ -28,4 +28,4 @@ _deps
 *.kdev4
 DartConfiguration.tcl
 bazel-*
-
+.cache