about summary refs log tree commit diff
path: root/third_party/abseil_cpp/absl/strings/internal/str_format/float_conversion.cc
diff options
context:
space:
mode:
Diffstat (limited to 'third_party/abseil_cpp/absl/strings/internal/str_format/float_conversion.cc')
-rw-r--r--third_party/abseil_cpp/absl/strings/internal/str_format/float_conversion.cc1419
1 files changed, 1419 insertions, 0 deletions
diff --git a/third_party/abseil_cpp/absl/strings/internal/str_format/float_conversion.cc b/third_party/abseil_cpp/absl/strings/internal/str_format/float_conversion.cc
new file mode 100644
index 000000000000..0ded0a66afa9
--- /dev/null
+++ b/third_party/abseil_cpp/absl/strings/internal/str_format/float_conversion.cc
@@ -0,0 +1,1419 @@
+// Copyright 2020 The Abseil Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//      https://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "absl/strings/internal/str_format/float_conversion.h"
+
+#include <string.h>
+
+#include <algorithm>
+#include <cassert>
+#include <cmath>
+#include <limits>
+#include <string>
+
+#include "absl/base/attributes.h"
+#include "absl/base/config.h"
+#include "absl/base/internal/bits.h"
+#include "absl/base/optimization.h"
+#include "absl/functional/function_ref.h"
+#include "absl/meta/type_traits.h"
+#include "absl/numeric/int128.h"
+#include "absl/strings/numbers.h"
+#include "absl/types/optional.h"
+#include "absl/types/span.h"
+
+namespace absl {
+ABSL_NAMESPACE_BEGIN
+namespace str_format_internal {
+
+namespace {
+
+// The code below wants to avoid heap allocations.
+// To do so it needs to allocate memory on the stack.
+// `StackArray` will allocate memory on the stack in the form of a uint32_t
+// array and call the provided callback with said memory.
+// It will allocate memory in increments of 512 bytes. We could allocate the
+// largest needed unconditionally, but that is more than we need in most of
+// cases. This way we use less stack in the common cases.
+class StackArray {
+  using Func = absl::FunctionRef<void(absl::Span<uint32_t>)>;
+  static constexpr size_t kStep = 512 / sizeof(uint32_t);
+  // 5 steps is 2560 bytes, which is enough to hold a long double with the
+  // largest/smallest exponents.
+  // The operations below will static_assert their particular maximum.
+  static constexpr size_t kNumSteps = 5;
+
+  // We do not want this function to be inlined.
+  // Otherwise the caller will allocate the stack space unnecessarily for all
+  // the variants even though it only calls one.
+  template <size_t steps>
+  ABSL_ATTRIBUTE_NOINLINE static void RunWithCapacityImpl(Func f) {
+    uint32_t values[steps * kStep]{};
+    f(absl::MakeSpan(values));
+  }
+
+ public:
+  static constexpr size_t kMaxCapacity = kStep * kNumSteps;
+
+  static void RunWithCapacity(size_t capacity, Func f) {
+    assert(capacity <= kMaxCapacity);
+    const size_t step = (capacity + kStep - 1) / kStep;
+    assert(step <= kNumSteps);
+    switch (step) {
+      case 1:
+        return RunWithCapacityImpl<1>(f);
+      case 2:
+        return RunWithCapacityImpl<2>(f);
+      case 3:
+        return RunWithCapacityImpl<3>(f);
+      case 4:
+        return RunWithCapacityImpl<4>(f);
+      case 5:
+        return RunWithCapacityImpl<5>(f);
+    }
+
+    assert(false && "Invalid capacity");
+  }
+};
+
+// Calculates `10 * (*v) + carry` and stores the result in `*v` and returns
+// the carry.
+template <typename Int>
+inline Int MultiplyBy10WithCarry(Int *v, Int carry) {
+  using BiggerInt = absl::conditional_t<sizeof(Int) == 4, uint64_t, uint128>;
+  BiggerInt tmp = 10 * static_cast<BiggerInt>(*v) + carry;
+  *v = static_cast<Int>(tmp);
+  return static_cast<Int>(tmp >> (sizeof(Int) * 8));
+}
+
+// Calculates `(2^64 * carry + *v) / 10`.
+// Stores the quotient in `*v` and returns the remainder.
+// Requires: `0 <= carry <= 9`
+inline uint64_t DivideBy10WithCarry(uint64_t *v, uint64_t carry) {
+  constexpr uint64_t divisor = 10;
+  // 2^64 / divisor = chunk_quotient + chunk_remainder / divisor
+  constexpr uint64_t chunk_quotient = (uint64_t{1} << 63) / (divisor / 2);
+  constexpr uint64_t chunk_remainder = uint64_t{} - chunk_quotient * divisor;
+
+  const uint64_t mod = *v % divisor;
+  const uint64_t next_carry = chunk_remainder * carry + mod;
+  *v = *v / divisor + carry * chunk_quotient + next_carry / divisor;
+  return next_carry % divisor;
+}
+
+// Generates the decimal representation for an integer of the form `v * 2^exp`,
+// where `v` and `exp` are both positive integers.
+// It generates the digits from the left (ie the most significant digit first)
+// to allow for direct printing into the sink.
+//
+// Requires `0 <= exp` and `exp <= numeric_limits<long double>::max_exponent`.
+class BinaryToDecimal {
+  static constexpr int ChunksNeeded(int exp) {
+    // We will left shift a uint128 by `exp` bits, so we need `128+exp` total
+    // bits. Round up to 32.
+    // See constructor for details about adding `10%` to the value.
+    return (128 + exp + 31) / 32 * 11 / 10;
+  }
+
+ public:
+  // Run the conversion for `v * 2^exp` and call `f(binary_to_decimal)`.
+  // This function will allocate enough stack space to perform the conversion.
+  static void RunConversion(uint128 v, int exp,
+                            absl::FunctionRef<void(BinaryToDecimal)> f) {
+    assert(exp > 0);
+    assert(exp <= std::numeric_limits<long double>::max_exponent);
+    static_assert(
+        static_cast<int>(StackArray::kMaxCapacity) >=
+            ChunksNeeded(std::numeric_limits<long double>::max_exponent),
+        "");
+
+    StackArray::RunWithCapacity(
+        ChunksNeeded(exp),
+        [=](absl::Span<uint32_t> input) { f(BinaryToDecimal(input, v, exp)); });
+  }
+
+  int TotalDigits() const {
+    return static_cast<int>((decimal_end_ - decimal_start_) * kDigitsPerChunk +
+                            CurrentDigits().size());
+  }
+
+  // See the current block of digits.
+  absl::string_view CurrentDigits() const {
+    return absl::string_view(digits_ + kDigitsPerChunk - size_, size_);
+  }
+
+  // Advance the current view of digits.
+  // Returns `false` when no more digits are available.
+  bool AdvanceDigits() {
+    if (decimal_start_ >= decimal_end_) return false;
+
+    uint32_t w = data_[decimal_start_++];
+    for (size_ = 0; size_ < kDigitsPerChunk; w /= 10) {
+      digits_[kDigitsPerChunk - ++size_] = w % 10 + '0';
+    }
+    return true;
+  }
+
+ private:
+  BinaryToDecimal(absl::Span<uint32_t> data, uint128 v, int exp) : data_(data) {
+    // We need to print the digits directly into the sink object without
+    // buffering them all first. To do this we need two things:
+    // - to know the total number of digits to do padding when necessary
+    // - to generate the decimal digits from the left.
+    //
+    // In order to do this, we do a two pass conversion.
+    // On the first pass we convert the binary representation of the value into
+    // a decimal representation in which each uint32_t chunk holds up to 9
+    // decimal digits.  In the second pass we take each decimal-holding-uint32_t
+    // value and generate the ascii decimal digits into `digits_`.
+    //
+    // The binary and decimal representations actually share the same memory
+    // region. As we go converting the chunks from binary to decimal we free
+    // them up and reuse them for the decimal representation. One caveat is that
+    // the decimal representation is around 7% less efficient in space than the
+    // binary one. We allocate an extra 10% memory to account for this. See
+    // ChunksNeeded for this calculation.
+    int chunk_index = exp / 32;
+    decimal_start_ = decimal_end_ = ChunksNeeded(exp);
+    const int offset = exp % 32;
+    // Left shift v by exp bits.
+    data_[chunk_index] = static_cast<uint32_t>(v << offset);
+    for (v >>= (32 - offset); v; v >>= 32)
+      data_[++chunk_index] = static_cast<uint32_t>(v);
+
+    while (chunk_index >= 0) {
+      // While we have more than one chunk available, go in steps of 1e9.
+      // `data_[chunk_index]` holds the highest non-zero binary chunk, so keep
+      // the variable updated.
+      uint32_t carry = 0;
+      for (int i = chunk_index; i >= 0; --i) {
+        uint64_t tmp = uint64_t{data_[i]} + (uint64_t{carry} << 32);
+        data_[i] = static_cast<uint32_t>(tmp / uint64_t{1000000000});
+        carry = static_cast<uint32_t>(tmp % uint64_t{1000000000});
+      }
+
+      // If the highest chunk is now empty, remove it from view.
+      if (data_[chunk_index] == 0) --chunk_index;
+
+      --decimal_start_;
+      assert(decimal_start_ != chunk_index);
+      data_[decimal_start_] = carry;
+    }
+
+    // Fill the first set of digits. The first chunk might not be complete, so
+    // handle differently.
+    for (uint32_t first = data_[decimal_start_++]; first != 0; first /= 10) {
+      digits_[kDigitsPerChunk - ++size_] = first % 10 + '0';
+    }
+  }
+
+ private:
+  static constexpr int kDigitsPerChunk = 9;
+
+  int decimal_start_;
+  int decimal_end_;
+
+  char digits_[kDigitsPerChunk];
+  int size_ = 0;
+
+  absl::Span<uint32_t> data_;
+};
+
+// Converts a value of the form `x * 2^-exp` into a sequence of decimal digits.
+// Requires `-exp < 0` and
+// `-exp >= limits<long double>::min_exponent - limits<long double>::digits`.
+class FractionalDigitGenerator {
+ public:
+  // Run the conversion for `v * 2^exp` and call `f(generator)`.
+  // This function will allocate enough stack space to perform the conversion.
+  static void RunConversion(
+      uint128 v, int exp, absl::FunctionRef<void(FractionalDigitGenerator)> f) {
+    using Limits = std::numeric_limits<long double>;
+    assert(-exp < 0);
+    assert(-exp >= Limits::min_exponent - 128);
+    static_assert(StackArray::kMaxCapacity >=
+                      (Limits::digits + 128 - Limits::min_exponent + 31) / 32,
+                  "");
+    StackArray::RunWithCapacity((Limits::digits + exp + 31) / 32,
+                                [=](absl::Span<uint32_t> input) {
+                                  f(FractionalDigitGenerator(input, v, exp));
+                                });
+  }
+
+  // Returns true if there are any more non-zero digits left.
+  bool HasMoreDigits() const { return next_digit_ != 0 || chunk_index_ >= 0; }
+
+  // Returns true if the remainder digits are greater than 5000...
+  bool IsGreaterThanHalf() const {
+    return next_digit_ > 5 || (next_digit_ == 5 && chunk_index_ >= 0);
+  }
+  // Returns true if the remainder digits are exactly 5000...
+  bool IsExactlyHalf() const { return next_digit_ == 5 && chunk_index_ < 0; }
+
+  struct Digits {
+    int digit_before_nine;
+    int num_nines;
+  };
+
+  // Get the next set of digits.
+  // They are composed by a non-9 digit followed by a runs of zero or more 9s.
+  Digits GetDigits() {
+    Digits digits{next_digit_, 0};
+
+    next_digit_ = GetOneDigit();
+    while (next_digit_ == 9) {
+      ++digits.num_nines;
+      next_digit_ = GetOneDigit();
+    }
+
+    return digits;
+  }
+
+ private:
+  // Return the next digit.
+  int GetOneDigit() {
+    if (chunk_index_ < 0) return 0;
+
+    uint32_t carry = 0;
+    for (int i = chunk_index_; i >= 0; --i) {
+      carry = MultiplyBy10WithCarry(&data_[i], carry);
+    }
+    // If the lowest chunk is now empty, remove it from view.
+    if (data_[chunk_index_] == 0) --chunk_index_;
+    return carry;
+  }
+
+  FractionalDigitGenerator(absl::Span<uint32_t> data, uint128 v, int exp)
+      : chunk_index_(exp / 32), data_(data) {
+    const int offset = exp % 32;
+    // Right shift `v` by `exp` bits.
+    data_[chunk_index_] = static_cast<uint32_t>(v << (32 - offset));
+    v >>= offset;
+    // Make sure we don't overflow the data. We already calculated that
+    // non-zero bits fit, so we might not have space for leading zero bits.
+    for (int pos = chunk_index_; v; v >>= 32)
+      data_[--pos] = static_cast<uint32_t>(v);
+
+    // Fill next_digit_, as GetDigits expects it to be populated always.
+    next_digit_ = GetOneDigit();
+  }
+
+  int next_digit_;
+  int chunk_index_;
+  absl::Span<uint32_t> data_;
+};
+
+// Count the number of leading zero bits.
+int LeadingZeros(uint64_t v) { return base_internal::CountLeadingZeros64(v); }
+int LeadingZeros(uint128 v) {
+  auto high = static_cast<uint64_t>(v >> 64);
+  auto low = static_cast<uint64_t>(v);
+  return high != 0 ? base_internal::CountLeadingZeros64(high)
+                   : 64 + base_internal::CountLeadingZeros64(low);
+}
+
+// Round up the text digits starting at `p`.
+// The buffer must have an extra digit that is known to not need rounding.
+// This is done below by having an extra '0' digit on the left.
+void RoundUp(char *p) {
+  while (*p == '9' || *p == '.') {
+    if (*p == '9') *p = '0';
+    --p;
+  }
+  ++*p;
+}
+
+// Check the previous digit and round up or down to follow the round-to-even
+// policy.
+void RoundToEven(char *p) {
+  if (*p == '.') --p;
+  if (*p % 2 == 1) RoundUp(p);
+}
+
+// Simple integral decimal digit printing for values that fit in 64-bits.
+// Returns the pointer to the last written digit.
+char *PrintIntegralDigitsFromRightFast(uint64_t v, char *p) {
+  do {
+    *--p = DivideBy10WithCarry(&v, 0) + '0';
+  } while (v != 0);
+  return p;
+}
+
+// Simple integral decimal digit printing for values that fit in 128-bits.
+// Returns the pointer to the last written digit.
+char *PrintIntegralDigitsFromRightFast(uint128 v, char *p) {
+  auto high = static_cast<uint64_t>(v >> 64);
+  auto low = static_cast<uint64_t>(v);
+
+  while (high != 0) {
+    uint64_t carry = DivideBy10WithCarry(&high, 0);
+    carry = DivideBy10WithCarry(&low, carry);
+    *--p = carry + '0';
+  }
+  return PrintIntegralDigitsFromRightFast(low, p);
+}
+
+// Simple fractional decimal digit printing for values that fir in 64-bits after
+// shifting.
+// Performs rounding if necessary to fit within `precision`.
+// Returns the pointer to one after the last character written.
+char *PrintFractionalDigitsFast(uint64_t v, char *start, int exp,
+                                int precision) {
+  char *p = start;
+  v <<= (64 - exp);
+  while (precision > 0) {
+    if (!v) return p;
+    *p++ = MultiplyBy10WithCarry(&v, uint64_t{0}) + '0';
+    --precision;
+  }
+
+  // We need to round.
+  if (v < 0x8000000000000000) {
+    // We round down, so nothing to do.
+  } else if (v > 0x8000000000000000) {
+    // We round up.
+    RoundUp(p - 1);
+  } else {
+    RoundToEven(p - 1);
+  }
+
+  assert(precision == 0);
+  // Precision can only be zero here.
+  return p;
+}
+
+// Simple fractional decimal digit printing for values that fir in 128-bits
+// after shifting.
+// Performs rounding if necessary to fit within `precision`.
+// Returns the pointer to one after the last character written.
+char *PrintFractionalDigitsFast(uint128 v, char *start, int exp,
+                                int precision) {
+  char *p = start;
+  v <<= (128 - exp);
+  auto high = static_cast<uint64_t>(v >> 64);
+  auto low = static_cast<uint64_t>(v);
+
+  // While we have digits to print and `low` is not empty, do the long
+  // multiplication.
+  while (precision > 0 && low != 0) {
+    uint64_t carry = MultiplyBy10WithCarry(&low, uint64_t{0});
+    carry = MultiplyBy10WithCarry(&high, carry);
+
+    *p++ = carry + '0';
+    --precision;
+  }
+
+  // Now `low` is empty, so use a faster approach for the rest of the digits.
+  // This block is pretty much the same as the main loop for the 64-bit case
+  // above.
+  while (precision > 0) {
+    if (!high) return p;
+    *p++ = MultiplyBy10WithCarry(&high, uint64_t{0}) + '0';
+    --precision;
+  }
+
+  // We need to round.
+  if (high < 0x8000000000000000) {
+    // We round down, so nothing to do.
+  } else if (high > 0x8000000000000000 || low != 0) {
+    // We round up.
+    RoundUp(p - 1);
+  } else {
+    RoundToEven(p - 1);
+  }
+
+  assert(precision == 0);
+  // Precision can only be zero here.
+  return p;
+}
+
+struct FormatState {
+  char sign_char;
+  int precision;
+  const FormatConversionSpecImpl &conv;
+  FormatSinkImpl *sink;
+
+  // In `alt` mode (flag #) we keep the `.` even if there are no fractional
+  // digits. In non-alt mode, we strip it.
+  bool ShouldPrintDot() const { return precision != 0 || conv.has_alt_flag(); }
+};
+
+struct Padding {
+  int left_spaces;
+  int zeros;
+  int right_spaces;
+};
+
+Padding ExtraWidthToPadding(size_t total_size, const FormatState &state) {
+  if (state.conv.width() < 0 ||
+      static_cast<size_t>(state.conv.width()) <= total_size) {
+    return {0, 0, 0};
+  }
+  int missing_chars = state.conv.width() - total_size;
+  if (state.conv.has_left_flag()) {
+    return {0, 0, missing_chars};
+  } else if (state.conv.has_zero_flag()) {
+    return {0, missing_chars, 0};
+  } else {
+    return {missing_chars, 0, 0};
+  }
+}
+
+void FinalPrint(const FormatState &state, absl::string_view data,
+                int padding_offset, int trailing_zeros,
+                absl::string_view data_postfix) {
+  if (state.conv.width() < 0) {
+    // No width specified. Fast-path.
+    if (state.sign_char != '\0') state.sink->Append(1, state.sign_char);
+    state.sink->Append(data);
+    state.sink->Append(trailing_zeros, '0');
+    state.sink->Append(data_postfix);
+    return;
+  }
+
+  auto padding = ExtraWidthToPadding((state.sign_char != '\0' ? 1 : 0) +
+                                         data.size() + data_postfix.size() +
+                                         static_cast<size_t>(trailing_zeros),
+                                     state);
+
+  state.sink->Append(padding.left_spaces, ' ');
+  if (state.sign_char != '\0') state.sink->Append(1, state.sign_char);
+  // Padding in general needs to be inserted somewhere in the middle of `data`.
+  state.sink->Append(data.substr(0, padding_offset));
+  state.sink->Append(padding.zeros, '0');
+  state.sink->Append(data.substr(padding_offset));
+  state.sink->Append(trailing_zeros, '0');
+  state.sink->Append(data_postfix);
+  state.sink->Append(padding.right_spaces, ' ');
+}
+
+// Fastpath %f formatter for when the shifted value fits in a simple integral
+// type.
+// Prints `v*2^exp` with the options from `state`.
+template <typename Int>
+void FormatFFast(Int v, int exp, const FormatState &state) {
+  constexpr int input_bits = sizeof(Int) * 8;
+
+  static constexpr size_t integral_size =
+      /* in case we need to round up an extra digit */ 1 +
+      /* decimal digits for uint128 */ 40 + 1;
+  char buffer[integral_size + /* . */ 1 + /* max digits uint128 */ 128];
+  buffer[integral_size] = '.';
+  char *const integral_digits_end = buffer + integral_size;
+  char *integral_digits_start;
+  char *const fractional_digits_start = buffer + integral_size + 1;
+  char *fractional_digits_end = fractional_digits_start;
+
+  if (exp >= 0) {
+    const int total_bits = input_bits - LeadingZeros(v) + exp;
+    integral_digits_start =
+        total_bits <= 64
+            ? PrintIntegralDigitsFromRightFast(static_cast<uint64_t>(v) << exp,
+                                               integral_digits_end)
+            : PrintIntegralDigitsFromRightFast(static_cast<uint128>(v) << exp,
+                                               integral_digits_end);
+  } else {
+    exp = -exp;
+
+    integral_digits_start = PrintIntegralDigitsFromRightFast(
+        exp < input_bits ? v >> exp : 0, integral_digits_end);
+    // PrintFractionalDigits may pull a carried 1 all the way up through the
+    // integral portion.
+    integral_digits_start[-1] = '0';
+
+    fractional_digits_end =
+        exp <= 64 ? PrintFractionalDigitsFast(v, fractional_digits_start, exp,
+                                              state.precision)
+                  : PrintFractionalDigitsFast(static_cast<uint128>(v),
+                                              fractional_digits_start, exp,
+                                              state.precision);
+    // There was a carry, so include the first digit too.
+    if (integral_digits_start[-1] != '0') --integral_digits_start;
+  }
+
+  size_t size = fractional_digits_end - integral_digits_start;
+
+  // In `alt` mode (flag #) we keep the `.` even if there are no fractional
+  // digits. In non-alt mode, we strip it.
+  if (!state.ShouldPrintDot()) --size;
+  FinalPrint(state, absl::string_view(integral_digits_start, size),
+             /*padding_offset=*/0,
+             static_cast<int>(state.precision - (fractional_digits_end -
+                                                 fractional_digits_start)),
+             /*data_postfix=*/"");
+}
+
+// Slow %f formatter for when the shifted value does not fit in a uint128, and
+// `exp > 0`.
+// Prints `v*2^exp` with the options from `state`.
+// This one is guaranteed to not have fractional digits, so we don't have to
+// worry about anything after the `.`.
+void FormatFPositiveExpSlow(uint128 v, int exp, const FormatState &state) {
+  BinaryToDecimal::RunConversion(v, exp, [&](BinaryToDecimal btd) {
+    const size_t total_digits =
+        btd.TotalDigits() +
+        (state.ShouldPrintDot() ? static_cast<size_t>(state.precision) + 1 : 0);
+
+    const auto padding = ExtraWidthToPadding(
+        total_digits + (state.sign_char != '\0' ? 1 : 0), state);
+
+    state.sink->Append(padding.left_spaces, ' ');
+    if (state.sign_char != '\0') state.sink->Append(1, state.sign_char);
+    state.sink->Append(padding.zeros, '0');
+
+    do {
+      state.sink->Append(btd.CurrentDigits());
+    } while (btd.AdvanceDigits());
+
+    if (state.ShouldPrintDot()) state.sink->Append(1, '.');
+    state.sink->Append(state.precision, '0');
+    state.sink->Append(padding.right_spaces, ' ');
+  });
+}
+
+// Slow %f formatter for when the shifted value does not fit in a uint128, and
+// `exp < 0`.
+// Prints `v*2^exp` with the options from `state`.
+// This one is guaranteed to be < 1.0, so we don't have to worry about integral
+// digits.
+void FormatFNegativeExpSlow(uint128 v, int exp, const FormatState &state) {
+  const size_t total_digits =
+      /* 0 */ 1 +
+      (state.ShouldPrintDot() ? static_cast<size_t>(state.precision) + 1 : 0);
+  auto padding =
+      ExtraWidthToPadding(total_digits + (state.sign_char ? 1 : 0), state);
+  padding.zeros += 1;
+  state.sink->Append(padding.left_spaces, ' ');
+  if (state.sign_char != '\0') state.sink->Append(1, state.sign_char);
+  state.sink->Append(padding.zeros, '0');
+
+  if (state.ShouldPrintDot()) state.sink->Append(1, '.');
+
+  // Print digits
+  int digits_to_go = state.precision;
+
+  FractionalDigitGenerator::RunConversion(
+      v, exp, [&](FractionalDigitGenerator digit_gen) {
+        // There are no digits to print here.
+        if (state.precision == 0) return;
+
+        // We go one digit at a time, while keeping track of runs of nines.
+        // The runs of nines are used to perform rounding when necessary.
+
+        while (digits_to_go > 0 && digit_gen.HasMoreDigits()) {
+          auto digits = digit_gen.GetDigits();
+
+          // Now we have a digit and a run of nines.
+          // See if we can print them all.
+          if (digits.num_nines + 1 < digits_to_go) {
+            // We don't have to round yet, so print them.
+            state.sink->Append(1, digits.digit_before_nine + '0');
+            state.sink->Append(digits.num_nines, '9');
+            digits_to_go -= digits.num_nines + 1;
+
+          } else {
+            // We can't print all the nines, see where we have to truncate.
+
+            bool round_up = false;
+            if (digits.num_nines + 1 > digits_to_go) {
+              // We round up at a nine. No need to print them.
+              round_up = true;
+            } else {
+              // We can fit all the nines, but truncate just after it.
+              if (digit_gen.IsGreaterThanHalf()) {
+                round_up = true;
+              } else if (digit_gen.IsExactlyHalf()) {
+                // Round to even
+                round_up =
+                    digits.num_nines != 0 || digits.digit_before_nine % 2 == 1;
+              }
+            }
+
+            if (round_up) {
+              state.sink->Append(1, digits.digit_before_nine + '1');
+              --digits_to_go;
+              // The rest will be zeros.
+            } else {
+              state.sink->Append(1, digits.digit_before_nine + '0');
+              state.sink->Append(digits_to_go - 1, '9');
+              digits_to_go = 0;
+            }
+            return;
+          }
+        }
+      });
+
+  state.sink->Append(digits_to_go, '0');
+  state.sink->Append(padding.right_spaces, ' ');
+}
+
+template <typename Int>
+void FormatF(Int mantissa, int exp, const FormatState &state) {
+  if (exp >= 0) {
+    const int total_bits = sizeof(Int) * 8 - LeadingZeros(mantissa) + exp;
+
+    // Fallback to the slow stack-based approach if we can't do it in a 64 or
+    // 128 bit state.
+    if (ABSL_PREDICT_FALSE(total_bits > 128)) {
+      return FormatFPositiveExpSlow(mantissa, exp, state);
+    }
+  } else {
+    // Fallback to the slow stack-based approach if we can't do it in a 64 or
+    // 128 bit state.
+    if (ABSL_PREDICT_FALSE(exp < -128)) {
+      return FormatFNegativeExpSlow(mantissa, -exp, state);
+    }
+  }
+  return FormatFFast(mantissa, exp, state);
+}
+
+// Grab the group of four bits (nibble) from `n`. E.g., nibble 1 corresponds to
+// bits 4-7.
+template <typename Int>
+uint8_t GetNibble(Int n, int nibble_index) {
+  constexpr Int mask_low_nibble = Int{0xf};
+  int shift = nibble_index * 4;
+  n &= mask_low_nibble << shift;
+  return static_cast<uint8_t>((n >> shift) & 0xf);
+}
+
+// Add one to the given nibble, applying carry to higher nibbles. Returns true
+// if overflow, false otherwise.
+template <typename Int>
+bool IncrementNibble(int nibble_index, Int *n) {
+  constexpr int kShift = sizeof(Int) * 8 - 1;
+  constexpr int kNumNibbles = sizeof(Int) * 8 / 4;
+  Int before = *n >> kShift;
+  // Here we essentially want to take the number 1 and move it into the requsted
+  // nibble, then add it to *n to effectively increment the nibble. However,
+  // ASan will complain if we try to shift the 1 beyond the limits of the Int,
+  // i.e., if the nibble_index is out of range. So therefore we check for this
+  // and if we are out of range we just add 0 which leaves *n unchanged, which
+  // seems like the reasonable thing to do in that case.
+  *n += ((nibble_index >= kNumNibbles) ? 0 : (Int{1} << (nibble_index * 4)));
+  Int after = *n >> kShift;
+  return (before && !after) || (nibble_index >= kNumNibbles);
+}
+
+// Return a mask with 1's in the given nibble and all lower nibbles.
+template <typename Int>
+Int MaskUpToNibbleInclusive(int nibble_index) {
+  constexpr int kNumNibbles = sizeof(Int) * 8 / 4;
+  static const Int ones = ~Int{0};
+  return ones >> std::max(0, 4 * (kNumNibbles - nibble_index - 1));
+}
+
+// Return a mask with 1's below the given nibble.
+template <typename Int>
+Int MaskUpToNibbleExclusive(int nibble_index) {
+  return nibble_index <= 0 ? 0 : MaskUpToNibbleInclusive<Int>(nibble_index - 1);
+}
+
+template <typename Int>
+Int MoveToNibble(uint8_t nibble, int nibble_index) {
+  return Int{nibble} << (4 * nibble_index);
+}
+
+// Given mantissa size, find optimal # of mantissa bits to put in initial digit.
+//
+// In the hex representation we keep a single hex digit to the left of the dot.
+// However, the question as to how many bits of the mantissa should be put into
+// that hex digit in theory is arbitrary, but in practice it is optimal to
+// choose based on the size of the mantissa. E.g., for a `double`, there are 53
+// mantissa bits, so that means that we should put 1 bit to the left of the dot,
+// thereby leaving 52 bits to the right, which is evenly divisible by four and
+// thus all fractional digits represent actual precision. For a `long double`,
+// on the other hand, there are 64 bits of mantissa, thus we can use all four
+// bits for the initial hex digit and still have a number left over (60) that is
+// a multiple of four. Once again, the goal is to have all fractional digits
+// represent real precision.
+template <typename Float>
+constexpr int HexFloatLeadingDigitSizeInBits() {
+  return std::numeric_limits<Float>::digits % 4 > 0
+             ? std::numeric_limits<Float>::digits % 4
+             : 4;
+}
+
+// This function captures the rounding behavior of glibc for hex float
+// representations. E.g. when rounding 0x1.ab800000 to a precision of .2
+// ("%.2a") glibc will round up because it rounds toward the even number (since
+// 0xb is an odd number, it will round up to 0xc). However, when rounding at a
+// point that is not followed by 800000..., it disregards the parity and rounds
+// up if > 8 and rounds down if < 8.
+template <typename Int>
+bool HexFloatNeedsRoundUp(Int mantissa, int final_nibble_displayed,
+                          uint8_t leading) {
+  // If the last nibble (hex digit) to be displayed is the lowest on in the
+  // mantissa then that means that we don't have any further nibbles to inform
+  // rounding, so don't round.
+  if (final_nibble_displayed <= 0) {
+    return false;
+  }
+  int rounding_nibble_idx = final_nibble_displayed - 1;
+  constexpr int kTotalNibbles = sizeof(Int) * 8 / 4;
+  assert(final_nibble_displayed <= kTotalNibbles);
+  Int mantissa_up_to_rounding_nibble_inclusive =
+      mantissa & MaskUpToNibbleInclusive<Int>(rounding_nibble_idx);
+  Int eight = MoveToNibble<Int>(8, rounding_nibble_idx);
+  if (mantissa_up_to_rounding_nibble_inclusive != eight) {
+    return mantissa_up_to_rounding_nibble_inclusive > eight;
+  }
+  // Nibble in question == 8.
+  uint8_t round_if_odd = (final_nibble_displayed == kTotalNibbles)
+                             ? leading
+                             : GetNibble(mantissa, final_nibble_displayed);
+  return round_if_odd % 2 == 1;
+}
+
+// Stores values associated with a Float type needed by the FormatA
+// implementation in order to avoid templatizing that function by the Float
+// type.
+struct HexFloatTypeParams {
+  template <typename Float>
+  explicit HexFloatTypeParams(Float)
+      : min_exponent(std::numeric_limits<Float>::min_exponent - 1),
+        leading_digit_size_bits(HexFloatLeadingDigitSizeInBits<Float>()) {
+    assert(leading_digit_size_bits >= 1 && leading_digit_size_bits <= 4);
+  }
+
+  int min_exponent;
+  int leading_digit_size_bits;
+};
+
+// Hex Float Rounding. First check if we need to round; if so, then we do that
+// by manipulating (incrementing) the mantissa, that way we can later print the
+// mantissa digits by iterating through them in the same way regardless of
+// whether a rounding happened.
+template <typename Int>
+void FormatARound(bool precision_specified, const FormatState &state,
+                  uint8_t *leading, Int *mantissa, int *exp) {
+  constexpr int kTotalNibbles = sizeof(Int) * 8 / 4;
+  // Index of the last nibble that we could display given precision.
+  int final_nibble_displayed =
+      precision_specified ? std::max(0, (kTotalNibbles - state.precision)) : 0;
+  if (HexFloatNeedsRoundUp(*mantissa, final_nibble_displayed, *leading)) {
+    // Need to round up.
+    bool overflow = IncrementNibble(final_nibble_displayed, mantissa);
+    *leading += (overflow ? 1 : 0);
+    if (ABSL_PREDICT_FALSE(*leading > 15)) {
+      // We have overflowed the leading digit. This would mean that we would
+      // need two hex digits to the left of the dot, which is not allowed. So
+      // adjust the mantissa and exponent so that the result is always 1.0eXXX.
+      *leading = 1;
+      *mantissa = 0;
+      *exp += 4;
+    }
+  }
+  // Now that we have handled a possible round-up we can go ahead and zero out
+  // all the nibbles of the mantissa that we won't need.
+  if (precision_specified) {
+    *mantissa &= ~MaskUpToNibbleExclusive<Int>(final_nibble_displayed);
+  }
+}
+
+template <typename Int>
+void FormatANormalize(const HexFloatTypeParams float_traits, uint8_t *leading,
+                      Int *mantissa, int *exp) {
+  constexpr int kIntBits = sizeof(Int) * 8;
+  static const Int kHighIntBit = Int{1} << (kIntBits - 1);
+  const int kLeadDigitBitsCount = float_traits.leading_digit_size_bits;
+  // Normalize mantissa so that highest bit set is in MSB position, unless we
+  // get interrupted by the exponent threshold.
+  while (*mantissa && !(*mantissa & kHighIntBit)) {
+    if (ABSL_PREDICT_FALSE(*exp - 1 < float_traits.min_exponent)) {
+      *mantissa >>= (float_traits.min_exponent - *exp);
+      *exp = float_traits.min_exponent;
+      return;
+    }
+    *mantissa <<= 1;
+    --*exp;
+  }
+  // Extract bits for leading digit then shift them away leaving the
+  // fractional part.
+  *leading =
+      static_cast<uint8_t>(*mantissa >> (kIntBits - kLeadDigitBitsCount));
+  *exp -= (*mantissa != 0) ? kLeadDigitBitsCount : *exp;
+  *mantissa <<= kLeadDigitBitsCount;
+}
+
+template <typename Int>
+void FormatA(const HexFloatTypeParams float_traits, Int mantissa, int exp,
+             bool uppercase, const FormatState &state) {
+  // Int properties.
+  constexpr int kIntBits = sizeof(Int) * 8;
+  constexpr int kTotalNibbles = sizeof(Int) * 8 / 4;
+  // Did the user specify a precision explicitly?
+  const bool precision_specified = state.conv.precision() >= 0;
+
+  // ========== Normalize/Denormalize ==========
+  exp += kIntBits;  // make all digits fractional digits.
+  // This holds the (up to four) bits of leading digit, i.e., the '1' in the
+  // number 0x1.e6fp+2. It's always > 0 unless number is zero or denormal.
+  uint8_t leading = 0;
+  FormatANormalize(float_traits, &leading, &mantissa, &exp);
+
+  // =============== Rounding ==================
+  // Check if we need to round; if so, then we do that by manipulating
+  // (incrementing) the mantissa before beginning to print characters.
+  FormatARound(precision_specified, state, &leading, &mantissa, &exp);
+
+  // ============= Format Result ===============
+  // This buffer holds the "0x1.ab1de3" portion of "0x1.ab1de3pe+2". Compute the
+  // size with long double which is the largest of the floats.
+  constexpr size_t kBufSizeForHexFloatRepr =
+      2                                               // 0x
+      + std::numeric_limits<long double>::digits / 4  // number of hex digits
+      + 1                                             // round up
+      + 1;                                            // "." (dot)
+  char digits_buffer[kBufSizeForHexFloatRepr];
+  char *digits_iter = digits_buffer;
+  const char *const digits =
+      static_cast<const char *>("0123456789ABCDEF0123456789abcdef") +
+      (uppercase ? 0 : 16);
+
+  // =============== Hex Prefix ================
+  *digits_iter++ = '0';
+  *digits_iter++ = uppercase ? 'X' : 'x';
+
+  // ========== Non-Fractional Digit ===========
+  *digits_iter++ = digits[leading];
+
+  // ================== Dot ====================
+  // There are three reasons we might need a dot. Keep in mind that, at this
+  // point, the mantissa holds only the fractional part.
+  if ((precision_specified && state.precision > 0) ||
+      (!precision_specified && mantissa > 0) || state.conv.has_alt_flag()) {
+    *digits_iter++ = '.';
+  }
+
+  // ============ Fractional Digits ============
+  int digits_emitted = 0;
+  while (mantissa > 0) {
+    *digits_iter++ = digits[GetNibble(mantissa, kTotalNibbles - 1)];
+    mantissa <<= 4;
+    ++digits_emitted;
+  }
+  int trailing_zeros =
+      precision_specified ? state.precision - digits_emitted : 0;
+  assert(trailing_zeros >= 0);
+  auto digits_result = string_view(digits_buffer, digits_iter - digits_buffer);
+
+  // =============== Exponent ==================
+  constexpr size_t kBufSizeForExpDecRepr =
+      numbers_internal::kFastToBufferSize  // requred for FastIntToBuffer
+      + 1                                  // 'p' or 'P'
+      + 1;                                 // '+' or '-'
+  char exp_buffer[kBufSizeForExpDecRepr];
+  exp_buffer[0] = uppercase ? 'P' : 'p';
+  exp_buffer[1] = exp >= 0 ? '+' : '-';
+  numbers_internal::FastIntToBuffer(exp < 0 ? -exp : exp, exp_buffer + 2);
+
+  // ============ Assemble Result ==============
+  FinalPrint(state,           //
+             digits_result,   // 0xN.NNN...
+             2,               // offset in `data` to start padding if needed.
+             trailing_zeros,  // num remaining mantissa padding zeros
+             exp_buffer);     // exponent
+}
+
+char *CopyStringTo(absl::string_view v, char *out) {
+  std::memcpy(out, v.data(), v.size());
+  return out + v.size();
+}
+
+template <typename Float>
+bool FallbackToSnprintf(const Float v, const FormatConversionSpecImpl &conv,
+                        FormatSinkImpl *sink) {
+  int w = conv.width() >= 0 ? conv.width() : 0;
+  int p = conv.precision() >= 0 ? conv.precision() : -1;
+  char fmt[32];
+  {
+    char *fp = fmt;
+    *fp++ = '%';
+    fp = CopyStringTo(FormatConversionSpecImplFriend::FlagsToString(conv), fp);
+    fp = CopyStringTo("*.*", fp);
+    if (std::is_same<long double, Float>()) {
+      *fp++ = 'L';
+    }
+    *fp++ = FormatConversionCharToChar(conv.conversion_char());
+    *fp = 0;
+    assert(fp < fmt + sizeof(fmt));
+  }
+  std::string space(512, '\0');
+  absl::string_view result;
+  while (true) {
+    int n = snprintf(&space[0], space.size(), fmt, w, p, v);
+    if (n < 0) return false;
+    if (static_cast<size_t>(n) < space.size()) {
+      result = absl::string_view(space.data(), n);
+      break;
+    }
+    space.resize(n + 1);
+  }
+  sink->Append(result);
+  return true;
+}
+
+// 128-bits in decimal: ceil(128*log(2)/log(10))
+//   or std::numeric_limits<__uint128_t>::digits10
+constexpr int kMaxFixedPrecision = 39;
+
+constexpr int kBufferLength = /*sign*/ 1 +
+                              /*integer*/ kMaxFixedPrecision +
+                              /*point*/ 1 +
+                              /*fraction*/ kMaxFixedPrecision +
+                              /*exponent e+123*/ 5;
+
+struct Buffer {
+  void push_front(char c) {
+    assert(begin > data);
+    *--begin = c;
+  }
+  void push_back(char c) {
+    assert(end < data + sizeof(data));
+    *end++ = c;
+  }
+  void pop_back() {
+    assert(begin < end);
+    --end;
+  }
+
+  char &back() {
+    assert(begin < end);
+    return end[-1];
+  }
+
+  char last_digit() const { return end[-1] == '.' ? end[-2] : end[-1]; }
+
+  int size() const { return static_cast<int>(end - begin); }
+
+  char data[kBufferLength];
+  char *begin;
+  char *end;
+};
+
+enum class FormatStyle { Fixed, Precision };
+
+// If the value is Inf or Nan, print it and return true.
+// Otherwise, return false.
+template <typename Float>
+bool ConvertNonNumericFloats(char sign_char, Float v,
+                             const FormatConversionSpecImpl &conv,
+                             FormatSinkImpl *sink) {
+  char text[4], *ptr = text;
+  if (sign_char != '\0') *ptr++ = sign_char;
+  if (std::isnan(v)) {
+    ptr = std::copy_n(
+        FormatConversionCharIsUpper(conv.conversion_char()) ? "NAN" : "nan", 3,
+        ptr);
+  } else if (std::isinf(v)) {
+    ptr = std::copy_n(
+        FormatConversionCharIsUpper(conv.conversion_char()) ? "INF" : "inf", 3,
+        ptr);
+  } else {
+    return false;
+  }
+
+  return sink->PutPaddedString(string_view(text, ptr - text), conv.width(), -1,
+                               conv.has_left_flag());
+}
+
+// Round up the last digit of the value.
+// It will carry over and potentially overflow. 'exp' will be adjusted in that
+// case.
+template <FormatStyle mode>
+void RoundUp(Buffer *buffer, int *exp) {
+  char *p = &buffer->back();
+  while (p >= buffer->begin && (*p == '9' || *p == '.')) {
+    if (*p == '9') *p = '0';
+    --p;
+  }
+
+  if (p < buffer->begin) {
+    *p = '1';
+    buffer->begin = p;
+    if (mode == FormatStyle::Precision) {
+      std::swap(p[1], p[2]);  // move the .
+      ++*exp;
+      buffer->pop_back();
+    }
+  } else {
+    ++*p;
+  }
+}
+
+void PrintExponent(int exp, char e, Buffer *out) {
+  out->push_back(e);
+  if (exp < 0) {
+    out->push_back('-');
+    exp = -exp;
+  } else {
+    out->push_back('+');
+  }
+  // Exponent digits.
+  if (exp > 99) {
+    out->push_back(exp / 100 + '0');
+    out->push_back(exp / 10 % 10 + '0');
+    out->push_back(exp % 10 + '0');
+  } else {
+    out->push_back(exp / 10 + '0');
+    out->push_back(exp % 10 + '0');
+  }
+}
+
+template <typename Float, typename Int>
+constexpr bool CanFitMantissa() {
+  return
+#if defined(__clang__) && !defined(__SSE3__)
+      // Workaround for clang bug: https://bugs.llvm.org/show_bug.cgi?id=38289
+      // Casting from long double to uint64_t is miscompiled and drops bits.
+      (!std::is_same<Float, long double>::value ||
+       !std::is_same<Int, uint64_t>::value) &&
+#endif
+      std::numeric_limits<Float>::digits <= std::numeric_limits<Int>::digits;
+}
+
+template <typename Float>
+struct Decomposed {
+  using MantissaType =
+      absl::conditional_t<std::is_same<long double, Float>::value, uint128,
+                          uint64_t>;
+  static_assert(std::numeric_limits<Float>::digits <= sizeof(MantissaType) * 8,
+                "");
+  MantissaType mantissa;
+  int exponent;
+};
+
+// Decompose the double into an integer mantissa and an exponent.
+template <typename Float>
+Decomposed<Float> Decompose(Float v) {
+  int exp;
+  Float m = std::frexp(v, &exp);
+  m = std::ldexp(m, std::numeric_limits<Float>::digits);
+  exp -= std::numeric_limits<Float>::digits;
+
+  return {static_cast<typename Decomposed<Float>::MantissaType>(m), exp};
+}
+
+// Print 'digits' as decimal.
+// In Fixed mode, we add a '.' at the end.
+// In Precision mode, we add a '.' after the first digit.
+template <FormatStyle mode, typename Int>
+int PrintIntegralDigits(Int digits, Buffer *out) {
+  int printed = 0;
+  if (digits) {
+    for (; digits; digits /= 10) out->push_front(digits % 10 + '0');
+    printed = out->size();
+    if (mode == FormatStyle::Precision) {
+      out->push_front(*out->begin);
+      out->begin[1] = '.';
+    } else {
+      out->push_back('.');
+    }
+  } else if (mode == FormatStyle::Fixed) {
+    out->push_front('0');
+    out->push_back('.');
+    printed = 1;
+  }
+  return printed;
+}
+
+// Back out 'extra_digits' digits and round up if necessary.
+bool RemoveExtraPrecision(int extra_digits, bool has_leftover_value,
+                          Buffer *out, int *exp_out) {
+  if (extra_digits <= 0) return false;
+
+  // Back out the extra digits
+  out->end -= extra_digits;
+
+  bool needs_to_round_up = [&] {
+    // We look at the digit just past the end.
+    // There must be 'extra_digits' extra valid digits after end.
+    if (*out->end > '5') return true;
+    if (*out->end < '5') return false;
+    if (has_leftover_value || std::any_of(out->end + 1, out->end + extra_digits,
+                                          [](char c) { return c != '0'; }))
+      return true;
+
+    // Ends in ...50*, round to even.
+    return out->last_digit() % 2 == 1;
+  }();
+
+  if (needs_to_round_up) {
+    RoundUp<FormatStyle::Precision>(out, exp_out);
+  }
+  return true;
+}
+
+// Print the value into the buffer.
+// This will not include the exponent, which will be returned in 'exp_out' for
+// Precision mode.
+template <typename Int, typename Float, FormatStyle mode>
+bool FloatToBufferImpl(Int int_mantissa, int exp, int precision, Buffer *out,
+                       int *exp_out) {
+  assert((CanFitMantissa<Float, Int>()));
+
+  const int int_bits = std::numeric_limits<Int>::digits;
+
+  // In precision mode, we start printing one char to the right because it will
+  // also include the '.'
+  // In fixed mode we put the dot afterwards on the right.
+  out->begin = out->end =
+      out->data + 1 + kMaxFixedPrecision + (mode == FormatStyle::Precision);
+
+  if (exp >= 0) {
+    if (std::numeric_limits<Float>::digits + exp > int_bits) {
+      // The value will overflow the Int
+      return false;
+    }
+    int digits_printed = PrintIntegralDigits<mode>(int_mantissa << exp, out);
+    int digits_to_zero_pad = precision;
+    if (mode == FormatStyle::Precision) {
+      *exp_out = digits_printed - 1;
+      digits_to_zero_pad -= digits_printed - 1;
+      if (RemoveExtraPrecision(-digits_to_zero_pad, false, out, exp_out)) {
+        return true;
+      }
+    }
+    for (; digits_to_zero_pad-- > 0;) out->push_back('0');
+    return true;
+  }
+
+  exp = -exp;
+  // We need at least 4 empty bits for the next decimal digit.
+  // We will multiply by 10.
+  if (exp > int_bits - 4) return false;
+
+  const Int mask = (Int{1} << exp) - 1;
+
+  // Print the integral part first.
+  int digits_printed = PrintIntegralDigits<mode>(int_mantissa >> exp, out);
+  int_mantissa &= mask;
+
+  int fractional_count = precision;
+  if (mode == FormatStyle::Precision) {
+    if (digits_printed == 0) {
+      // Find the first non-zero digit, when in Precision mode.
+      *exp_out = 0;
+      if (int_mantissa) {
+        while (int_mantissa <= mask) {
+          int_mantissa *= 10;
+          --*exp_out;
+        }
+      }
+      out->push_front(static_cast<char>(int_mantissa >> exp) + '0');
+      out->push_back('.');
+      int_mantissa &= mask;
+    } else {
+      // We already have a digit, and a '.'
+      *exp_out = digits_printed - 1;
+      fractional_count -= *exp_out;
+      if (RemoveExtraPrecision(-fractional_count, int_mantissa != 0, out,
+                               exp_out)) {
+        // If we had enough digits, return right away.
+        // The code below will try to round again otherwise.
+        return true;
+      }
+    }
+  }
+
+  auto get_next_digit = [&] {
+    int_mantissa *= 10;
+    int digit = static_cast<int>(int_mantissa >> exp);
+    int_mantissa &= mask;
+    return digit;
+  };
+
+  // Print fractional_count more digits, if available.
+  for (; fractional_count > 0; --fractional_count) {
+    out->push_back(get_next_digit() + '0');
+  }
+
+  int next_digit = get_next_digit();
+  if (next_digit > 5 ||
+      (next_digit == 5 && (int_mantissa || out->last_digit() % 2 == 1))) {
+    RoundUp<mode>(out, exp_out);
+  }
+
+  return true;
+}
+
+template <FormatStyle mode, typename Float>
+bool FloatToBuffer(Decomposed<Float> decomposed, int precision, Buffer *out,
+                   int *exp) {
+  if (precision > kMaxFixedPrecision) return false;
+
+  // Try with uint64_t.
+  if (CanFitMantissa<Float, std::uint64_t>() &&
+      FloatToBufferImpl<std::uint64_t, Float, mode>(
+          static_cast<std::uint64_t>(decomposed.mantissa),
+          static_cast<std::uint64_t>(decomposed.exponent), precision, out, exp))
+    return true;
+
+#if defined(ABSL_HAVE_INTRINSIC_INT128)
+  // If that is not enough, try with __uint128_t.
+  return CanFitMantissa<Float, __uint128_t>() &&
+         FloatToBufferImpl<__uint128_t, Float, mode>(
+             static_cast<__uint128_t>(decomposed.mantissa),
+             static_cast<__uint128_t>(decomposed.exponent), precision, out,
+             exp);
+#endif
+  return false;
+}
+
+void WriteBufferToSink(char sign_char, absl::string_view str,
+                       const FormatConversionSpecImpl &conv,
+                       FormatSinkImpl *sink) {
+  int left_spaces = 0, zeros = 0, right_spaces = 0;
+  int missing_chars =
+      conv.width() >= 0 ? std::max(conv.width() - static_cast<int>(str.size()) -
+                                       static_cast<int>(sign_char != 0),
+                                   0)
+                        : 0;
+  if (conv.has_left_flag()) {
+    right_spaces = missing_chars;
+  } else if (conv.has_zero_flag()) {
+    zeros = missing_chars;
+  } else {
+    left_spaces = missing_chars;
+  }
+
+  sink->Append(left_spaces, ' ');
+  if (sign_char != '\0') sink->Append(1, sign_char);
+  sink->Append(zeros, '0');
+  sink->Append(str);
+  sink->Append(right_spaces, ' ');
+}
+
+template <typename Float>
+bool FloatToSink(const Float v, const FormatConversionSpecImpl &conv,
+                 FormatSinkImpl *sink) {
+  // Print the sign or the sign column.
+  Float abs_v = v;
+  char sign_char = 0;
+  if (std::signbit(abs_v)) {
+    sign_char = '-';
+    abs_v = -abs_v;
+  } else if (conv.has_show_pos_flag()) {
+    sign_char = '+';
+  } else if (conv.has_sign_col_flag()) {
+    sign_char = ' ';
+  }
+
+  // Print nan/inf.
+  if (ConvertNonNumericFloats(sign_char, abs_v, conv, sink)) {
+    return true;
+  }
+
+  int precision = conv.precision() < 0 ? 6 : conv.precision();
+
+  int exp = 0;
+
+  auto decomposed = Decompose(abs_v);
+
+  Buffer buffer;
+
+  FormatConversionChar c = conv.conversion_char();
+
+  if (c == FormatConversionCharInternal::f ||
+      c == FormatConversionCharInternal::F) {
+    FormatF(decomposed.mantissa, decomposed.exponent,
+            {sign_char, precision, conv, sink});
+    return true;
+  } else if (c == FormatConversionCharInternal::e ||
+             c == FormatConversionCharInternal::E) {
+    if (!FloatToBuffer<FormatStyle::Precision>(decomposed, precision, &buffer,
+                                               &exp)) {
+      return FallbackToSnprintf(v, conv, sink);
+    }
+    if (!conv.has_alt_flag() && buffer.back() == '.') buffer.pop_back();
+    PrintExponent(
+        exp, FormatConversionCharIsUpper(conv.conversion_char()) ? 'E' : 'e',
+        &buffer);
+  } else if (c == FormatConversionCharInternal::g ||
+             c == FormatConversionCharInternal::G) {
+    precision = std::max(0, precision - 1);
+    if (!FloatToBuffer<FormatStyle::Precision>(decomposed, precision, &buffer,
+                                               &exp)) {
+      return FallbackToSnprintf(v, conv, sink);
+    }
+    if (precision + 1 > exp && exp >= -4) {
+      if (exp < 0) {
+        // Have 1.23456, needs 0.00123456
+        // Move the first digit
+        buffer.begin[1] = *buffer.begin;
+        // Add some zeros
+        for (; exp < -1; ++exp) *buffer.begin-- = '0';
+        *buffer.begin-- = '.';
+        *buffer.begin = '0';
+      } else if (exp > 0) {
+        // Have 1.23456, needs 1234.56
+        // Move the '.' exp positions to the right.
+        std::rotate(buffer.begin + 1, buffer.begin + 2, buffer.begin + exp + 2);
+      }
+      exp = 0;
+    }
+    if (!conv.has_alt_flag()) {
+      while (buffer.back() == '0') buffer.pop_back();
+      if (buffer.back() == '.') buffer.pop_back();
+    }
+    if (exp) {
+      PrintExponent(
+          exp, FormatConversionCharIsUpper(conv.conversion_char()) ? 'E' : 'e',
+          &buffer);
+    }
+  } else if (c == FormatConversionCharInternal::a ||
+             c == FormatConversionCharInternal::A) {
+    bool uppercase = (c == FormatConversionCharInternal::A);
+    FormatA(HexFloatTypeParams(Float{}), decomposed.mantissa,
+            decomposed.exponent, uppercase, {sign_char, precision, conv, sink});
+    return true;
+  } else {
+    return false;
+  }
+
+  WriteBufferToSink(sign_char,
+                    absl::string_view(buffer.begin, buffer.end - buffer.begin),
+                    conv, sink);
+
+  return true;
+}
+
+}  // namespace
+
+bool ConvertFloatImpl(long double v, const FormatConversionSpecImpl &conv,
+                      FormatSinkImpl *sink) {
+  if (std::numeric_limits<long double>::digits ==
+      2 * std::numeric_limits<double>::digits) {
+    // This is the `double-double` representation of `long double`.
+    // We do not handle it natively. Fallback to snprintf.
+    return FallbackToSnprintf(v, conv, sink);
+  }
+
+  return FloatToSink(v, conv, sink);
+}
+
+bool ConvertFloatImpl(float v, const FormatConversionSpecImpl &conv,
+                      FormatSinkImpl *sink) {
+  return FloatToSink(static_cast<double>(v), conv, sink);
+}
+
+bool ConvertFloatImpl(double v, const FormatConversionSpecImpl &conv,
+                      FormatSinkImpl *sink) {
+  return FloatToSink(v, conv, sink);
+}
+
+}  // namespace str_format_internal
+ABSL_NAMESPACE_END
+}  // namespace absl