xref: /llvm-project/flang/runtime/edit-output.cpp (revision fc97d2e68b03bc2979395e84b645e5b3ba35aecd)

//===-- runtime/edit-output.cpp -------------------------------------------===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//

#include "edit-output.h"
#include "emit-encoded.h"
#include "utf.h"
#include "flang/Common/real.h"
#include "flang/Common/uint128.h"
#include <algorithm>

namespace Fortran::runtime::io {
RT_OFFLOAD_API_GROUP_BEGIN

// In output statement, add a space between numbers and characters.
static RT_API_ATTRS void addSpaceBeforeCharacter(IoStatementState &io) {
  if (auto *list{io.get_if<ListDirectedStatementState<Direction::Output>>()}) {
    list->set_lastWasUndelimitedCharacter(false);
  }
}

// B/O/Z output of arbitrarily sized data emits a binary/octal/hexadecimal
// representation of what is interpreted to be a single unsigned integer value.
// When used with character data, endianness is exposed.
template <int LOG2_BASE>
static RT_API_ATTRS bool EditBOZOutput(IoStatementState &io,
    const DataEdit &edit, const unsigned char *data0, std::size_t bytes) {
  addSpaceBeforeCharacter(io);
  int digits{static_cast<int>((bytes * 8) / LOG2_BASE)};
  int get{static_cast<int>(bytes * 8) - digits * LOG2_BASE};
  if (get > 0) {
    ++digits;
  } else {
    get = LOG2_BASE;
  }
  int shift{7};
  int increment{isHostLittleEndian ? -1 : 1};
  const unsigned char *data{data0 + (isHostLittleEndian ? bytes - 1 : 0)};
  int skippedZeroes{0};
  int digit{0};
  // The same algorithm is used to generate digits for real (below)
  // as well as for generating them only to skip leading zeroes (here).
  // Bits are copied one at a time from the source data.
  // TODO: Multiple bit copies for hexadecimal, where misalignment
  // is not possible; or for octal when all 3 bits come from the
  // same byte.
  while (bytes > 0) {
    if (get == 0) {
      if (digit != 0) {
        break; // first nonzero leading digit
      }
      ++skippedZeroes;
      get = LOG2_BASE;
    } else if (shift < 0) {
      data += increment;
      --bytes;
      shift = 7;
    } else {
      digit = 2 * digit + ((*data >> shift--) & 1);
      --get;
    }
  }
  // Emit leading spaces and zeroes; detect field overflow
  int leadingZeroes{0};
  int editWidth{edit.width.value_or(0)};
  int significant{digits - skippedZeroes};
  if (edit.digits && significant <= *edit.digits) { // Bw.m, Ow.m, Zw.m
    if (*edit.digits == 0 && bytes == 0) {
      editWidth = std::max(1, editWidth);
    } else {
      leadingZeroes = *edit.digits - significant;
    }
  } else if (bytes == 0) {
    leadingZeroes = 1;
  }
  int subTotal{leadingZeroes + significant};
  int leadingSpaces{std::max(0, editWidth - subTotal)};
  if (editWidth > 0 && leadingSpaces + subTotal > editWidth) {
    return EmitRepeated(io, '*', editWidth);
  }
  if (!(EmitRepeated(io, ' ', leadingSpaces) &&
          EmitRepeated(io, '0', leadingZeroes))) {
    return false;
  }
  // Emit remaining digits
  while (bytes > 0) {
    if (get == 0) {
      char ch{static_cast<char>(digit >= 10 ? 'A' + digit - 10 : '0' + digit)};
      if (!EmitAscii(io, &ch, 1)) {
        return false;
      }
      get = LOG2_BASE;
      digit = 0;
    } else if (shift < 0) {
      data += increment;
      --bytes;
      shift = 7;
    } else {
      digit = 2 * digit + ((*data >> shift--) & 1);
      --get;
    }
  }
  return true;
}

template <int KIND>
bool RT_API_ATTRS EditIntegerOutput(IoStatementState &io, const DataEdit &edit,
    common::HostSignedIntType<8 * KIND> n, bool isSigned) {
  addSpaceBeforeCharacter(io);
  char buffer[130], *end{&buffer[sizeof buffer]}, *p{end};
  bool isNegative{isSigned && n < 0};
  using Unsigned = common::HostUnsignedIntType<8 * KIND>;
  Unsigned un{static_cast<Unsigned>(n)};
  int signChars{0};
  switch (edit.descriptor) {
  case DataEdit::ListDirected:
  case 'G':
  case 'I':
    if (isNegative) {
      un = -un;
    }
    if (isNegative || (edit.modes.editingFlags & signPlus)) {
      signChars = 1; // '-' or '+'
    }
    while (un > 0) {
      auto quotient{un / 10u};
      *--p = '0' + static_cast<int>(un - Unsigned{10} * quotient);
      un = quotient;
    }
    break;
  case 'B':
    return EditBOZOutput<1>(
        io, edit, reinterpret_cast<const unsigned char *>(&n), KIND);
  case 'O':
    return EditBOZOutput<3>(
        io, edit, reinterpret_cast<const unsigned char *>(&n), KIND);
  case 'Z':
    return EditBOZOutput<4>(
        io, edit, reinterpret_cast<const unsigned char *>(&n), KIND);
  case 'L':
    return EditLogicalOutput(io, edit, n != 0 ? true : false);
  case 'A': // legacy extension
    return EditCharacterOutput(
        io, edit, reinterpret_cast<char *>(&n), sizeof n);
  default:
    io.GetIoErrorHandler().SignalError(IostatErrorInFormat,
        "Data edit descriptor '%c' may not be used with an INTEGER data item",
        edit.descriptor);
    return false;
  }

  int digits = end - p;
  int leadingZeroes{0};
  int editWidth{edit.width.value_or(0)};
  if (edit.descriptor == 'I' && edit.digits && digits <= *edit.digits) {
    // Only Iw.m can produce leading zeroes, not Gw.d (F'202X 13.7.5.2.2)
    if (*edit.digits == 0 && n == 0) {
      // Iw.0 with zero value: output field must be blank.  For I0.0
      // and a zero value, emit one blank character.
      signChars = 0; // in case of SP
      editWidth = std::max(1, editWidth);
    } else {
      leadingZeroes = *edit.digits - digits;
    }
  } else if (n == 0) {
    leadingZeroes = 1;
  }
  int subTotal{signChars + leadingZeroes + digits};
  int leadingSpaces{std::max(0, editWidth - subTotal)};
  if (editWidth > 0 && leadingSpaces + subTotal > editWidth) {
    return EmitRepeated(io, '*', editWidth);
  }
  if (edit.IsListDirected()) {
    int total{std::max(leadingSpaces, 1) + subTotal};
    if (io.GetConnectionState().NeedAdvance(static_cast<std::size_t>(total)) &&
        !io.AdvanceRecord()) {
      return false;
    }
    leadingSpaces = 1;
  }
  return EmitRepeated(io, ' ', leadingSpaces) &&
      EmitAscii(io, n < 0 ? "-" : "+", signChars) &&
      EmitRepeated(io, '0', leadingZeroes) && EmitAscii(io, p, digits);
}

// Formats the exponent (see table 13.1 for all the cases)
RT_API_ATTRS const char *RealOutputEditingBase::FormatExponent(
    int expo, const DataEdit &edit, int &length) {
  char *eEnd{&exponent_[sizeof exponent_]};
  char *exponent{eEnd};
  for (unsigned e{static_cast<unsigned>(std::abs(expo))}; e > 0;) {
    unsigned quotient{e / 10u};
    *--exponent = '0' + e - 10 * quotient;
    e = quotient;
  }
  bool overflow{false};
  if (edit.expoDigits) {
    if (int ed{*edit.expoDigits}) { // Ew.dEe with e > 0
      overflow = exponent + ed < eEnd;
      while (exponent > exponent_ + 2 /*E+*/ && exponent + ed > eEnd) {
        *--exponent = '0';
      }
    } else if (exponent == eEnd) {
      *--exponent = '0'; // Ew.dE0 with zero-valued exponent
    }
  } else if (edit.variation == 'X') {
    if (expo == 0) {
      *--exponent = '0'; // EX without Ee and zero-valued exponent
    }
  } else {
    // Ensure at least two exponent digits unless EX
    while (exponent + 2 > eEnd) {
      *--exponent = '0';
    }
  }
  *--exponent = expo < 0 ? '-' : '+';
  if (edit.variation == 'X') {
    *--exponent = 'P';
  } else if (edit.expoDigits || edit.IsListDirected() || exponent + 3 == eEnd) {
    *--exponent = edit.descriptor == 'D' ? 'D' : 'E'; // not 'G' or 'Q'
  }
  length = eEnd - exponent;
  return overflow ? nullptr : exponent;
}

RT_API_ATTRS bool RealOutputEditingBase::EmitPrefix(
    const DataEdit &edit, std::size_t length, std::size_t width) {
  if (edit.IsListDirected()) {
    int prefixLength{edit.descriptor == DataEdit::ListDirectedRealPart ? 2
            : edit.descriptor == DataEdit::ListDirectedImaginaryPart   ? 0
                                                                       : 1};
    int suffixLength{edit.descriptor == DataEdit::ListDirectedRealPart ||
                edit.descriptor == DataEdit::ListDirectedImaginaryPart
            ? 1
            : 0};
    length += prefixLength + suffixLength;
    ConnectionState &connection{io_.GetConnectionState()};
    return (!connection.NeedAdvance(length) || io_.AdvanceRecord()) &&
        EmitAscii(io_, " (", prefixLength);
  } else if (width > length) {
    return EmitRepeated(io_, ' ', width - length);
  } else {
    return true;
  }
}

RT_API_ATTRS bool RealOutputEditingBase::EmitSuffix(const DataEdit &edit) {
  if (edit.descriptor == DataEdit::ListDirectedRealPart) {
    return EmitAscii(
        io_, edit.modes.editingFlags & decimalComma ? ";" : ",", 1);
  } else if (edit.descriptor == DataEdit::ListDirectedImaginaryPart) {
    return EmitAscii(io_, ")", 1);
  } else {
    return true;
  }
}

template <int KIND>
RT_API_ATTRS decimal::ConversionToDecimalResult
RealOutputEditing<KIND>::ConvertToDecimal(
    int significantDigits, enum decimal::FortranRounding rounding, int flags) {
  auto converted{decimal::ConvertToDecimal<binaryPrecision>(buffer_,
      sizeof buffer_, static_cast<enum decimal::DecimalConversionFlags>(flags),
      significantDigits, rounding, x_)};
  if (!converted.str) { // overflow
    io_.GetIoErrorHandler().Crash(
        "RealOutputEditing::ConvertToDecimal: buffer size %zd was insufficient",
        sizeof buffer_);
  }
  return converted;
}

static RT_API_ATTRS bool IsInfOrNaN(const char *p, int length) {
  if (!p || length < 1) {
    return false;
  }
  if (*p == '-' || *p == '+') {
    if (length == 1) {
      return false;
    }
    ++p;
  }
  return *p == 'I' || *p == 'N';
}

// 13.7.2.3.3 in F'2018
template <int KIND>
RT_API_ATTRS bool RealOutputEditing<KIND>::EditEorDOutput(
    const DataEdit &edit) {
  addSpaceBeforeCharacter(io_);
  int editDigits{edit.digits.value_or(0)}; // 'd' field
  int editWidth{edit.width.value_or(0)}; // 'w' field
  int significantDigits{editDigits};
  int flags{0};
  if (edit.modes.editingFlags & signPlus) {
    flags |= decimal::AlwaysSign;
  }
  int scale{edit.modes.scale}; // 'kP' value
  bool isEN{edit.variation == 'N'};
  bool isES{edit.variation == 'S'};
  if (editWidth == 0) { // "the processor selects the field width"
    if (edit.digits.has_value()) { // E0.d
      if (editDigits == 0 && scale <= 0) { // E0.0
        significantDigits = isEN || isES ? 0 : 1;
      }
    } else { // E0
      flags |= decimal::Minimize;
      significantDigits =
          sizeof buffer_ - 5; // sign, NUL, + 3 extra for EN scaling
    }
  }
  int zeroesAfterPoint{0};
  if (isEN) {
    scale = IsZero() ? 1 : 3;
    significantDigits += scale;
  } else if (isES) {
    scale = 1;
    ++significantDigits;
  } else if (scale < 0) {
    if (scale <= -editDigits) {
      io_.GetIoErrorHandler().SignalError(IostatBadScaleFactor,
          "Scale factor (kP) %d cannot be less than -d (%d)", scale,
          -editDigits);
      return false;
    }
    zeroesAfterPoint = -scale;
    significantDigits = std::max(0, significantDigits - zeroesAfterPoint);
  } else if (scale > 0) {
    if (scale >= editDigits + 2) {
      io_.GetIoErrorHandler().SignalError(IostatBadScaleFactor,
          "Scale factor (kP) %d cannot be greater than d+2 (%d)", scale,
          editDigits + 2);
      return false;
    }
    ++significantDigits;
    scale = std::min(scale, significantDigits + 1);
  } else if (edit.digits.value_or(1) == 0 && !edit.variation) {
    // F'2023 13.7.2.3.3 p5; does not apply to Gw.0(Ee) or E0(no d)
    io_.GetIoErrorHandler().SignalError(IostatErrorInFormat,
        "Output edit descriptor %cw.d must have d>0", edit.descriptor);
    return false;
  }
  // In EN editing, multiple attempts may be necessary, so this is a loop.
  while (true) {
    decimal::ConversionToDecimalResult converted{
        ConvertToDecimal(significantDigits, edit.modes.round, flags)};
    if (IsInfOrNaN(converted.str, static_cast<int>(converted.length))) {
      return editWidth > 0 &&
              converted.length + trailingBlanks_ >
                  static_cast<std::size_t>(editWidth)
          ? EmitRepeated(io_, '*', editWidth)
          : EmitPrefix(edit, converted.length, editWidth) &&
              EmitAscii(io_, converted.str, converted.length) &&
              EmitRepeated(io_, ' ', trailingBlanks_) && EmitSuffix(edit);
    }
    if (!IsZero()) {
      converted.decimalExponent -= scale;
    }
    if (isEN) {
      // EN mode: we need an effective exponent field that is
      // a multiple of three.
      if (int modulus{converted.decimalExponent % 3}; modulus != 0) {
        if (significantDigits > 1) {
          --significantDigits;
          --scale;
          continue;
        }
        // Rounded nines up to a 1.
        scale += modulus;
        converted.decimalExponent -= modulus;
      }
      if (scale > 3) {
        int adjust{3 * (scale / 3)};
        scale -= adjust;
        converted.decimalExponent += adjust;
      } else if (scale < 1) {
        int adjust{3 - 3 * (scale / 3)};
        scale += adjust;
        converted.decimalExponent -= adjust;
      }
      significantDigits = editDigits + scale;
    }
    // Format the exponent (see table 13.1 for all the cases)
    int expoLength{0};
    const char *exponent{
        FormatExponent(converted.decimalExponent, edit, expoLength)};
    int signLength{*converted.str == '-' || *converted.str == '+' ? 1 : 0};
    int convertedDigits{static_cast<int>(converted.length) - signLength};
    int zeroesBeforePoint{std::max(0, scale - convertedDigits)};
    int digitsBeforePoint{std::max(0, scale - zeroesBeforePoint)};
    int digitsAfterPoint{convertedDigits - digitsBeforePoint};
    int trailingZeroes{flags & decimal::Minimize
            ? 0
            : std::max(0,
                  significantDigits - (convertedDigits + zeroesBeforePoint))};
    int totalLength{signLength + digitsBeforePoint + zeroesBeforePoint +
        1 /*'.'*/ + zeroesAfterPoint + digitsAfterPoint + trailingZeroes +
        expoLength};
    int width{editWidth > 0 ? editWidth : totalLength};
    if (totalLength > width || !exponent) {
      return EmitRepeated(io_, '*', width);
    }
    if (totalLength < width && digitsBeforePoint == 0 &&
        zeroesBeforePoint == 0) {
      zeroesBeforePoint = 1;
      ++totalLength;
    }
    if (totalLength < width && editWidth == 0) {
      width = totalLength;
    }
    return EmitPrefix(edit, totalLength, width) &&
        EmitAscii(io_, converted.str, signLength + digitsBeforePoint) &&
        EmitRepeated(io_, '0', zeroesBeforePoint) &&
        EmitAscii(io_, edit.modes.editingFlags & decimalComma ? "," : ".", 1) &&
        EmitRepeated(io_, '0', zeroesAfterPoint) &&
        EmitAscii(io_, converted.str + signLength + digitsBeforePoint,
            digitsAfterPoint) &&
        EmitRepeated(io_, '0', trailingZeroes) &&
        EmitAscii(io_, exponent, expoLength) && EmitSuffix(edit);
  }
}

// 13.7.2.3.2 in F'2018
template <int KIND>
RT_API_ATTRS bool RealOutputEditing<KIND>::EditFOutput(const DataEdit &edit) {
  addSpaceBeforeCharacter(io_);
  int fracDigits{edit.digits.value_or(0)}; // 'd' field
  const int editWidth{edit.width.value_or(0)}; // 'w' field
  enum decimal::FortranRounding rounding{edit.modes.round};
  int flags{0};
  if (edit.modes.editingFlags & signPlus) {
    flags |= decimal::AlwaysSign;
  }
  if (editWidth == 0) { // "the processor selects the field width"
    if (!edit.digits.has_value()) { // F0
      flags |= decimal::Minimize;
      fracDigits = sizeof buffer_ - 2; // sign & NUL
    }
  }
  bool emitTrailingZeroes{!(flags & decimal::Minimize)};
  // Multiple conversions may be needed to get the right number of
  // effective rounded fractional digits.
  bool canIncrease{true};
  for (int extraDigits{fracDigits == 0 ? 1 : 0};;) {
    decimal::ConversionToDecimalResult converted{
        ConvertToDecimal(extraDigits + fracDigits, rounding, flags)};
    const char *convertedStr{converted.str};
    if (IsInfOrNaN(convertedStr, static_cast<int>(converted.length))) {
      return editWidth > 0 &&
              converted.length > static_cast<std::size_t>(editWidth)
          ? EmitRepeated(io_, '*', editWidth)
          : EmitPrefix(edit, converted.length, editWidth) &&
              EmitAscii(io_, convertedStr, converted.length) &&
              EmitSuffix(edit);
    }
    int expo{converted.decimalExponent + edit.modes.scale /*kP*/};
    int signLength{*convertedStr == '-' || *convertedStr == '+' ? 1 : 0};
    int convertedDigits{static_cast<int>(converted.length) - signLength};
    if (IsZero()) { // don't treat converted "0" as significant digit
      expo = 0;
      convertedDigits = 0;
    }
    bool isNegative{*convertedStr == '-'};
    char one[2];
    if (expo > extraDigits && extraDigits >= 0 && canIncrease) {
      extraDigits = expo;
      if (!edit.digits.has_value()) { // F0
        fracDigits = sizeof buffer_ - extraDigits - 2; // sign & NUL
      }
      canIncrease = false; // only once
      continue;
    } else if (expo == -fracDigits && convertedDigits > 0) {
      // Result will be either a signed zero or power of ten, depending
      // on rounding.
      char leading{convertedStr[signLength]};
      bool roundToPowerOfTen{false};
      switch (edit.modes.round) {
      case decimal::FortranRounding::RoundUp:
        roundToPowerOfTen = !isNegative;
        break;
      case decimal::FortranRounding::RoundDown:
        roundToPowerOfTen = isNegative;
        break;
      case decimal::FortranRounding::RoundToZero:
        break;
      case decimal::FortranRounding::RoundNearest:
        if (leading == '5' &&
            rounding == decimal::FortranRounding::RoundNearest) {
          // Try again, rounding away from zero.
          rounding = isNegative ? decimal::FortranRounding::RoundDown
                                : decimal::FortranRounding::RoundUp;
          extraDigits = 1 - fracDigits; // just one digit needed
          continue;
        }
        roundToPowerOfTen = leading > '5';
        break;
      case decimal::FortranRounding::RoundCompatible:
        roundToPowerOfTen = leading >= '5';
        break;
      }
      if (roundToPowerOfTen) {
        ++expo;
        convertedDigits = 1;
        if (signLength > 0) {
          one[0] = *convertedStr;
          one[1] = '1';
        } else {
          one[0] = '1';
        }
        convertedStr = one;
      } else {
        expo = 0;
        convertedDigits = 0;
      }
    } else if (expo < extraDigits && extraDigits > -fracDigits) {
      extraDigits = std::max(expo, -fracDigits);
      continue;
    }
    int digitsBeforePoint{std::max(0, std::min(expo, convertedDigits))};
    int zeroesBeforePoint{std::max(0, expo - digitsBeforePoint)};
    if (zeroesBeforePoint > 0 && (flags & decimal::Minimize)) {
      // If a minimized result looks like an integer, emit all of
      // its digits rather than clipping some to zeroes.
      // This can happen with HUGE(0._2) == 65504._2.
      flags &= ~decimal::Minimize;
      continue;
    }
    int zeroesAfterPoint{std::min(fracDigits, std::max(0, -expo))};
    int digitsAfterPoint{convertedDigits - digitsBeforePoint};
    int trailingZeroes{emitTrailingZeroes
            ? std::max(0, fracDigits - (zeroesAfterPoint + digitsAfterPoint))
            : 0};
    if (digitsBeforePoint + zeroesBeforePoint + zeroesAfterPoint +
            digitsAfterPoint + trailingZeroes ==
        0) {
      zeroesBeforePoint = 1; // "." -> "0."
    }
    int totalLength{signLength + digitsBeforePoint + zeroesBeforePoint +
        1 /*'.'*/ + zeroesAfterPoint + digitsAfterPoint + trailingZeroes +
        trailingBlanks_ /* G editing converted to F */};
    int width{editWidth > 0 || trailingBlanks_ ? editWidth : totalLength};
    if (totalLength > width) {
      return EmitRepeated(io_, '*', width);
    }
    if (totalLength < width && digitsBeforePoint + zeroesBeforePoint == 0) {
      zeroesBeforePoint = 1;
      ++totalLength;
    }
    return EmitPrefix(edit, totalLength, width) &&
        EmitAscii(io_, convertedStr, signLength + digitsBeforePoint) &&
        EmitRepeated(io_, '0', zeroesBeforePoint) &&
        EmitAscii(io_, edit.modes.editingFlags & decimalComma ? "," : ".", 1) &&
        EmitRepeated(io_, '0', zeroesAfterPoint) &&
        EmitAscii(io_, convertedStr + signLength + digitsBeforePoint,
            digitsAfterPoint) &&
        EmitRepeated(io_, '0', trailingZeroes) &&
        EmitRepeated(io_, ' ', trailingBlanks_) && EmitSuffix(edit);
  }
}

// 13.7.5.2.3 in F'2018
template <int KIND>
RT_API_ATTRS DataEdit RealOutputEditing<KIND>::EditForGOutput(DataEdit edit) {
  edit.descriptor = 'E';
  edit.variation = 'G'; // to suppress error for Ew.0
  int editWidth{edit.width.value_or(0)};
  int significantDigits{edit.digits.value_or(
      static_cast<int>(BinaryFloatingPoint::decimalPrecision))}; // 'd'
  if (editWidth > 0 && significantDigits == 0) {
    return edit; // Gw.0Ee -> Ew.0Ee for w > 0
  }
  int flags{0};
  if (edit.modes.editingFlags & signPlus) {
    flags |= decimal::AlwaysSign;
  }
  decimal::ConversionToDecimalResult converted{
      ConvertToDecimal(significantDigits, edit.modes.round, flags)};
  if (IsInfOrNaN(converted.str, static_cast<int>(converted.length))) {
    return edit; // Inf/Nan -> Ew.d (same as Fw.d)
  }
  int expo{IsZero() ? 1 : converted.decimalExponent}; // 's'
  if (expo < 0 || expo > significantDigits) {
    if (editWidth == 0 && !edit.expoDigits) { // G0.d -> G0.dE0
      edit.expoDigits = 0;
    }
    return edit; // Ew.dEe
  }
  edit.descriptor = 'F';
  edit.modes.scale = 0; // kP is ignored for G when no exponent field
  trailingBlanks_ = 0;
  if (editWidth > 0) {
    int expoDigits{edit.expoDigits.value_or(0)};
    // F'2023 13.7.5.2.3 p5: "If 0 <= s <= d, the scale factor has no effect
    // and F(w − n).(d − s),n(’b’) editing is used where b is a blank and
    // n is 4 for Gw.d editing, e + 2 for Gw.dEe editing if e > 0, and
    // 4 for Gw.dE0 editing."
    trailingBlanks_ = expoDigits > 0 ? expoDigits + 2 : 4; // 'n'
  }
  if (edit.digits.has_value()) {
    *edit.digits = std::max(0, *edit.digits - expo);
  }
  return edit;
}

// 13.10.4 in F'2018
template <int KIND>
RT_API_ATTRS bool RealOutputEditing<KIND>::EditListDirectedOutput(
    const DataEdit &edit) {
  decimal::ConversionToDecimalResult converted{
      ConvertToDecimal(1, edit.modes.round)};
  if (IsInfOrNaN(converted.str, static_cast<int>(converted.length))) {
    DataEdit copy{edit};
    copy.variation = DataEdit::ListDirected;
    return EditEorDOutput(copy);
  }
  int expo{converted.decimalExponent};
  // The decimal precision of 16-bit floating-point types is very low,
  // so use a reasonable cap of 6 to allow more values to be emitted
  // with Fw.d editing.
  static constexpr int maxExpo{
      std::max(6, BinaryFloatingPoint::decimalPrecision)};
  if (expo < 0 || expo > maxExpo) {
    DataEdit copy{edit};
    copy.variation = DataEdit::ListDirected;
    copy.modes.scale = 1; // 1P
    return EditEorDOutput(copy);
  } else {
    return EditFOutput(edit);
  }
}

// 13.7.2.3.6 in F'2023
// The specification for hexadecimal output, unfortunately for implementors,
// leaves as "implementation dependent" the choice of how to emit values
// with multiple hexadecimal output possibilities that are numerically
// equivalent.  The one working implementation of EX output that I can find
// apparently chooses to frame the nybbles from most to least significant,
// rather than trying to minimize the magnitude of the binary exponent.
// E.g., 2. is edited into 0X8.0P-2 rather than 0X2.0P0.  This implementation
// follows that precedent so as to avoid a gratuitous incompatibility.
template <int KIND>
RT_API_ATTRS auto RealOutputEditing<KIND>::ConvertToHexadecimal(
    int significantDigits, enum decimal::FortranRounding rounding,
    int flags) -> ConvertToHexadecimalResult {
  if (x_.IsNaN() || x_.IsInfinite()) {
    auto converted{ConvertToDecimal(significantDigits, rounding, flags)};
    return {converted.str, static_cast<int>(converted.length), 0};
  }
  x_.RoundToBits(4 * significantDigits, rounding);
  if (x_.IsInfinite()) { // rounded away to +/-Inf
    auto converted{ConvertToDecimal(significantDigits, rounding, flags)};
    return {converted.str, static_cast<int>(converted.length), 0};
  }
  int len{0};
  if (x_.IsNegative()) {
    buffer_[len++] = '-';
  } else if (flags & decimal::AlwaysSign) {
    buffer_[len++] = '+';
  }
  auto fraction{x_.Fraction()};
  if (fraction == 0) {
    buffer_[len++] = '0';
    return {buffer_, len, 0};
  } else {
    // Ensure that the MSB is set.
    int expo{x_.UnbiasedExponent() - 3};
    while (!(fraction >> (x_.binaryPrecision - 1))) {
      fraction <<= 1;
      --expo;
    }
    // This is initially the right shift count needed to bring the
    // most-significant hexadecimal digit's bits into the LSBs.
    // x_.binaryPrecision is constant, so / can be used for readability.
    int shift{x_.binaryPrecision - 4};
    typename BinaryFloatingPoint::RawType one{1};
    auto remaining{(one << x_.binaryPrecision) - one};
    for (int digits{0}; digits < significantDigits; ++digits) {
      if ((flags & decimal::Minimize) && !(fraction & remaining)) {
        break;
      }
      int hexDigit{0};
      if (shift >= 0) {
        hexDigit = int(fraction >> shift) & 0xf;
      } else if (shift >= -3) {
        hexDigit = int(fraction << -shift) & 0xf;
      }
      if (hexDigit >= 10) {
        buffer_[len++] = 'A' + hexDigit - 10;
      } else {
        buffer_[len++] = '0' + hexDigit;
      }
      shift -= 4;
      remaining >>= 4;
    }
    return {buffer_, len, expo};
  }
}

template <int KIND>
RT_API_ATTRS bool RealOutputEditing<KIND>::EditEXOutput(const DataEdit &edit) {
  addSpaceBeforeCharacter(io_);
  int editDigits{edit.digits.value_or(0)}; // 'd' field
  int significantDigits{editDigits + 1};
  int flags{0};
  if (edit.modes.editingFlags & signPlus) {
    flags |= decimal::AlwaysSign;
  }
  int editWidth{edit.width.value_or(0)}; // 'w' field
  if ((editWidth == 0 && !edit.digits) || editDigits == 0) {
    // EX0 or EXw.0
    flags |= decimal::Minimize;
    static constexpr int maxSigHexDigits{
        (common::PrecisionOfRealKind(16) + 3) / 4};
    significantDigits = maxSigHexDigits;
  }
  auto converted{
      ConvertToHexadecimal(significantDigits, edit.modes.round, flags)};
  if (IsInfOrNaN(converted.str, converted.length)) {
    return editWidth > 0 && converted.length > editWidth
        ? EmitRepeated(io_, '*', editWidth)
        : (editWidth <= converted.length ||
              EmitRepeated(io_, ' ', editWidth - converted.length)) &&
            EmitAscii(io_, converted.str, converted.length);
  }
  int signLength{converted.length > 0 &&
              (converted.str[0] == '-' || converted.str[0] == '+')
          ? 1
          : 0};
  int convertedDigits{converted.length - signLength};
  int expoLength{0};
  const char *exponent{FormatExponent(converted.exponent, edit, expoLength)};
  int trailingZeroes{flags & decimal::Minimize
          ? 0
          : std::max(0, significantDigits - convertedDigits)};
  int totalLength{converted.length + trailingZeroes + expoLength + 3 /*0X.*/};
  int width{editWidth > 0 ? editWidth : totalLength};
  return totalLength > width || !exponent
      ? EmitRepeated(io_, '*', width)
      : EmitRepeated(io_, ' ', width - totalLength) &&
          EmitAscii(io_, converted.str, signLength) &&
          EmitAscii(io_, "0X", 2) &&
          EmitAscii(io_, converted.str + signLength, 1) &&
          EmitAscii(
              io_, edit.modes.editingFlags & decimalComma ? "," : ".", 1) &&
          EmitAscii(io_, converted.str + signLength + 1,
              converted.length - (signLength + 1)) &&
          EmitRepeated(io_, '0', trailingZeroes) &&
          EmitAscii(io_, exponent, expoLength);
}

template <int KIND>
RT_API_ATTRS bool RealOutputEditing<KIND>::Edit(const DataEdit &edit) {
  const DataEdit *editPtr{&edit};
  DataEdit newEdit;
  if (editPtr->descriptor == 'G') {
    // Avoid recursive call as in Edit(EditForGOutput(edit)).
    newEdit = EditForGOutput(*editPtr);
    editPtr = &newEdit;
    RUNTIME_CHECK(io_.GetIoErrorHandler(), editPtr->descriptor != 'G');
  }
  switch (editPtr->descriptor) {
  case 'D':
    return EditEorDOutput(*editPtr);
  case 'E':
    if (editPtr->variation == 'X') {
      return EditEXOutput(*editPtr);
    } else {
      return EditEorDOutput(*editPtr);
    }
  case 'F':
    return EditFOutput(*editPtr);
  case 'B':
    return EditBOZOutput<1>(io_, *editPtr,
        reinterpret_cast<const unsigned char *>(&x_),
        common::BitsForBinaryPrecision(common::PrecisionOfRealKind(KIND)) >> 3);
  case 'O':
    return EditBOZOutput<3>(io_, *editPtr,
        reinterpret_cast<const unsigned char *>(&x_),
        common::BitsForBinaryPrecision(common::PrecisionOfRealKind(KIND)) >> 3);
  case 'Z':
    return EditBOZOutput<4>(io_, *editPtr,
        reinterpret_cast<const unsigned char *>(&x_),
        common::BitsForBinaryPrecision(common::PrecisionOfRealKind(KIND)) >> 3);
  case 'L':
    return EditLogicalOutput(
        io_, *editPtr, *reinterpret_cast<const char *>(&x_));
  case 'A': // legacy extension
    return EditCharacterOutput(
        io_, *editPtr, reinterpret_cast<char *>(&x_), sizeof x_);
  default:
    if (editPtr->IsListDirected()) {
      return EditListDirectedOutput(*editPtr);
    }
    io_.GetIoErrorHandler().SignalError(IostatErrorInFormat,
        "Data edit descriptor '%c' may not be used with a REAL data item",
        editPtr->descriptor);
    return false;
  }
  return false;
}

RT_API_ATTRS bool ListDirectedLogicalOutput(IoStatementState &io,
    ListDirectedStatementState<Direction::Output> &list, bool truth) {
  return list.EmitLeadingSpaceOrAdvance(io) &&
      EmitAscii(io, truth ? "T" : "F", 1);
}

RT_API_ATTRS bool EditLogicalOutput(
    IoStatementState &io, const DataEdit &edit, bool truth) {
  switch (edit.descriptor) {
  case 'L':
  case 'G':
    return EmitRepeated(io, ' ', std::max(0, edit.width.value_or(1) - 1)) &&
        EmitAscii(io, truth ? "T" : "F", 1);
  case 'B':
    return EditBOZOutput<1>(io, edit,
        reinterpret_cast<const unsigned char *>(&truth), sizeof truth);
  case 'O':
    return EditBOZOutput<3>(io, edit,
        reinterpret_cast<const unsigned char *>(&truth), sizeof truth);
  case 'Z':
    return EditBOZOutput<4>(io, edit,
        reinterpret_cast<const unsigned char *>(&truth), sizeof truth);
  case 'A': { // legacy extension
    int truthBits{truth};
    int len{sizeof truthBits};
    int width{edit.width.value_or(len)};
    return EmitRepeated(io, ' ', std::max(0, width - len)) &&
        EmitEncoded(
            io, reinterpret_cast<char *>(&truthBits), std::min(width, len));
  }
  default:
    io.GetIoErrorHandler().SignalError(IostatErrorInFormat,
        "Data edit descriptor '%c' may not be used with a LOGICAL data item",
        edit.descriptor);
    return false;
  }
}

template <typename CHAR>
RT_API_ATTRS bool ListDirectedCharacterOutput(IoStatementState &io,
    ListDirectedStatementState<Direction::Output> &list, const CHAR *x,
    std::size_t length) {
  bool ok{true};
  MutableModes &modes{io.mutableModes()};
  ConnectionState &connection{io.GetConnectionState()};
  if (modes.delim) {
    ok = ok && list.EmitLeadingSpaceOrAdvance(io);
    // Value is delimited with ' or " marks, and interior
    // instances of that character are doubled.
    auto EmitOne{[&](CHAR ch) {
      if (connection.NeedAdvance(1)) {
        ok = ok && io.AdvanceRecord();
      }
      ok = ok && EmitEncoded(io, &ch, 1);
    }};
    EmitOne(modes.delim);
    for (std::size_t j{0}; j < length; ++j) {
      // Doubled delimiters must be put on the same record
      // in order to be acceptable as list-directed or NAMELIST
      // input; however, this requirement is not always possible
      // when the records have a fixed length, as is the case with
      // internal output.  The standard is silent on what should
      // happen, and no two extant Fortran implementations do
      // the same thing when tested with this case.
      // This runtime splits the doubled delimiters across
      // two records for lack of a better alternative.
      if (x[j] == static_cast<CHAR>(modes.delim)) {
        EmitOne(x[j]);
      }
      EmitOne(x[j]);
    }
    EmitOne(modes.delim);
  } else {
    // Undelimited list-directed output
    ok = ok && list.EmitLeadingSpaceOrAdvance(io, length > 0 ? 1 : 0, true);
    std::size_t put{0};
    std::size_t oneAtATime{
        connection.useUTF8<CHAR>() || connection.internalIoCharKind > 1
            ? 1
            : length};
    while (ok && put < length) {
      if (std::size_t chunk{std::min<std::size_t>(
              std::min<std::size_t>(length - put, oneAtATime),
              connection.RemainingSpaceInRecord())}) {
        ok = EmitEncoded(io, x + put, chunk);
        put += chunk;
      } else {
        ok = io.AdvanceRecord() && EmitAscii(io, " ", 1);
      }
    }
    list.set_lastWasUndelimitedCharacter(true);
  }
  return ok;
}

template <typename CHAR>
RT_API_ATTRS bool EditCharacterOutput(IoStatementState &io,
    const DataEdit &edit, const CHAR *x, std::size_t length) {
  int len{static_cast<int>(length)};
  int width{edit.width.value_or(len)};
  switch (edit.descriptor) {
  case 'A':
    break;
  case 'G':
    if (width == 0) {
      width = len;
    }
    break;
  case 'B':
    return EditBOZOutput<1>(io, edit,
        reinterpret_cast<const unsigned char *>(x), sizeof(CHAR) * length);
  case 'O':
    return EditBOZOutput<3>(io, edit,
        reinterpret_cast<const unsigned char *>(x), sizeof(CHAR) * length);
  case 'Z':
    return EditBOZOutput<4>(io, edit,
        reinterpret_cast<const unsigned char *>(x), sizeof(CHAR) * length);
  case 'L':
    return EditLogicalOutput(io, edit, *reinterpret_cast<const char *>(x));
  default:
    io.GetIoErrorHandler().SignalError(IostatErrorInFormat,
        "Data edit descriptor '%c' may not be used with a CHARACTER data item",
        edit.descriptor);
    return false;
  }
  return EmitRepeated(io, ' ', std::max(0, width - len)) &&
      EmitEncoded(io, x, std::min(width, len));
}

template RT_API_ATTRS bool EditIntegerOutput<1>(
    IoStatementState &, const DataEdit &, std::int8_t, bool);
template RT_API_ATTRS bool EditIntegerOutput<2>(
    IoStatementState &, const DataEdit &, std::int16_t, bool);
template RT_API_ATTRS bool EditIntegerOutput<4>(
    IoStatementState &, const DataEdit &, std::int32_t, bool);
template RT_API_ATTRS bool EditIntegerOutput<8>(
    IoStatementState &, const DataEdit &, std::int64_t, bool);
template RT_API_ATTRS bool EditIntegerOutput<16>(
    IoStatementState &, const DataEdit &, common::int128_t, bool);

template class RealOutputEditing<2>;
template class RealOutputEditing<3>;
template class RealOutputEditing<4>;
template class RealOutputEditing<8>;
template class RealOutputEditing<10>;
// TODO: double/double
template class RealOutputEditing<16>;

template RT_API_ATTRS bool ListDirectedCharacterOutput(IoStatementState &,
    ListDirectedStatementState<Direction::Output> &, const char *,
    std::size_t chars);
template RT_API_ATTRS bool ListDirectedCharacterOutput(IoStatementState &,
    ListDirectedStatementState<Direction::Output> &, const char16_t *,
    std::size_t chars);
template RT_API_ATTRS bool ListDirectedCharacterOutput(IoStatementState &,
    ListDirectedStatementState<Direction::Output> &, const char32_t *,
    std::size_t chars);

template RT_API_ATTRS bool EditCharacterOutput(
    IoStatementState &, const DataEdit &, const char *, std::size_t chars);
template RT_API_ATTRS bool EditCharacterOutput(
    IoStatementState &, const DataEdit &, const char16_t *, std::size_t chars);
template RT_API_ATTRS bool EditCharacterOutput(
    IoStatementState &, const DataEdit &, const char32_t *, std::size_t chars);

RT_OFFLOAD_API_GROUP_END
} // namespace Fortran::runtime::io