diff options
| author |  Abseil Team <absl-team@google.com> | 2020-09-18 15:55:15 -0700 |
|---|---|---|
| committer |  Derek Mauro <dmauro@google.com> | 2020-09-24 13:47:15 -0400 |
| commit | b56cbdd23834a65682c0b46f367f8679e83bc894 (patch) | |
| tree | dacab9a64dd1a9e9668737e511d1a5420ff96001 /absl/strings/internal/str_format | |
| parent | b832dce8489ef7b6231384909fd9b68d5a5ff2b7 (diff) | |
Abseil LTS 2020092320200923
What's New:
* `absl::StatusOr<T>` has been released. See our [blog
post](https://abseil.io/blog/2020-091021-status) for more
information.
* Abseil Flags reflection interfaces have been released.
* Abseil Flags memory usage has been significantly optimized.
* Abseil now supports a "hardened" build mode. This build mode enables
runtime checks that guard against programming errors that may lead
to security vulnerabilities.
Notable Fixes:
* Sanitizer dynamic annotations like `AnnotateRWLockCreate` that are
also defined by the compiler sanitizer implementation are no longer
also defined by Abseil.
* Sanitizer macros are now prefixed with `ABSL_` to avoid naming collisions.
* Sanitizer usage is now automatically detected and no longer requires
macros like `ADDRESS_SANITIZER` to be defined on the command line.
Breaking Changes:
* Abseil no longer contains a `dynamic_annotations` library. Users
using a supported build system (Bazel or CMake) are unaffected by
this, but users manually specifying link libraries may get an error
about a missing linker input.
Baseline: 7680a5f8efe32de4753baadbd63e74e59d95bac1
Cherry picks: None
Diffstat (limited to 'absl/strings/internal/str_format')
18 files changed, 2289 insertions, 570 deletions
diff --git a/absl/strings/internal/str_format/arg.cc b/absl/strings/internal/str_format/arg.cc index 4d0604e0..9feb2248 100644 --- a/absl/strings/internal/str_format/arg.cc +++ b/absl/strings/internal/str_format/arg.cc @@ -12,14 +12,13 @@ #include "absl/base/port.h" #include "absl/strings/internal/str_format/float_conversion.h" +#include "absl/strings/numbers.h" namespace absl { ABSL_NAMESPACE_BEGIN namespace str_format_internal { namespace { -const char kDigit[2][32] = { "0123456789abcdef", "0123456789ABCDEF" }; - // Reduce *capacity by s.size(), clipped to a 0 minimum. void ReducePadding(string_view s, size_t *capacity) { *capacity = Excess(s.size(), *capacity); @@ -48,125 +47,179 @@ struct IsSigned<absl::int128> : std::true_type {}; template <> struct IsSigned<absl::uint128> : std::false_type {}; -class ConvertedIntInfo { +// Integral digit printer. +// Call one of the PrintAs* routines after construction once. +// Use with_neg_and_zero/without_neg_or_zero/is_negative to access the results. +class IntDigits { public: + // Print the unsigned integer as octal. + // Supports unsigned integral types and uint128. + template <typename T> + void PrintAsOct(T v) { + static_assert(!IsSigned<T>::value, ""); + char *p = storage_ + sizeof(storage_); + do { + *--p = static_cast<char>('0' + (static_cast<size_t>(v) & 7)); + v >>= 3; + } while (v); + start_ = p; + size_ = storage_ + sizeof(storage_) - p; + } + + // Print the signed or unsigned integer as decimal. + // Supports all integral types. template <typename T> - ConvertedIntInfo(T v, ConversionChar conv) { - using Unsigned = typename MakeUnsigned<T>::type; - auto u = static_cast<Unsigned>(v); - if (IsNeg(v)) { - is_neg_ = true; - u = Unsigned{} - u; - } else { - is_neg_ = false; + void PrintAsDec(T v) { + static_assert(std::is_integral<T>::value, ""); + start_ = storage_; + size_ = numbers_internal::FastIntToBuffer(v, storage_) - storage_; + } + + void PrintAsDec(int128 v) { + auto u = static_cast<uint128>(v); + bool add_neg = false; + if (v < 0) { + add_neg = true; + u = uint128{} - u; } - UnsignedToStringRight(u, conv); + PrintAsDec(u, add_neg); } - string_view digits() const { - return {end() - size_, static_cast<size_t>(size_)}; + void PrintAsDec(uint128 v, bool add_neg = false) { + // This function can be sped up if needed. We can call FastIntToBuffer + // twice, or fix FastIntToBuffer to support uint128. + char *p = storage_ + sizeof(storage_); + do { + p -= 2; + numbers_internal::PutTwoDigits(static_cast<size_t>(v % 100), p); + v /= 100; + } while (v); + if (p[0] == '0') { + // We printed one too many hexits. + ++p; + } + if (add_neg) { + *--p = '-'; + } + size_ = storage_ + sizeof(storage_) - p; + start_ = p; } - bool is_neg() const { return is_neg_; } - private: - template <typename T, bool IsSigned> - struct IsNegImpl { - static bool Eval(T v) { return v < 0; } - }; + // Print the unsigned integer as hex using lowercase. + // Supports unsigned integral types and uint128. template <typename T> - struct IsNegImpl<T, false> { - static bool Eval(T) { - return false; + void PrintAsHexLower(T v) { + static_assert(!IsSigned<T>::value, ""); + char *p = storage_ + sizeof(storage_); + + do { + p -= 2; + constexpr const char* table = numbers_internal::kHexTable; + std::memcpy(p, table + 2 * (static_cast<size_t>(v) & 0xFF), 2); + if (sizeof(T) == 1) break; + v >>= 8; + } while (v); + if (p[0] == '0') { + // We printed one too many digits. + ++p; } - }; + start_ = p; + size_ = storage_ + sizeof(storage_) - p; + } + // Print the unsigned integer as hex using uppercase. + // Supports unsigned integral types and uint128. template <typename T> - bool IsNeg(T v) { - return IsNegImpl<T, IsSigned<T>::value>::Eval(v); + void PrintAsHexUpper(T v) { + static_assert(!IsSigned<T>::value, ""); + char *p = storage_ + sizeof(storage_); + + // kHexTable is only lowercase, so do it manually for uppercase. + do { + *--p = "0123456789ABCDEF"[static_cast<size_t>(v) & 15]; + v >>= 4; + } while (v); + start_ = p; + size_ = storage_ + sizeof(storage_) - p; } - template <typename T> - void UnsignedToStringRight(T u, ConversionChar conv) { - char *p = end(); - switch (FormatConversionCharRadix(conv)) { - default: - case 10: - for (; u; u /= 10) - *--p = static_cast<char>('0' + static_cast<size_t>(u % 10)); - break; - case 8: - for (; u; u /= 8) - *--p = static_cast<char>('0' + static_cast<size_t>(u % 8)); - break; - case 16: { - const char *digits = kDigit[FormatConversionCharIsUpper(conv) ? 1 : 0]; - for (; u; u /= 16) *--p = digits[static_cast<size_t>(u % 16)]; - break; - } - } - size_ = static_cast<int>(end() - p); + // The printed value including the '-' sign if available. + // For inputs of value `0`, this will return "0" + string_view with_neg_and_zero() const { return {start_, size_}; } + + // The printed value not including the '-' sign. + // For inputs of value `0`, this will return "". + string_view without_neg_or_zero() const { + static_assert('-' < '0', "The check below verifies both."); + size_t advance = start_[0] <= '0' ? 1 : 0; + return {start_ + advance, size_ - advance}; } - const char *end() const { return storage_ + sizeof(storage_); } - char *end() { return storage_ + sizeof(storage_); } + bool is_negative() const { return start_[0] == '-'; } - bool is_neg_; - int size_; - // Max size: 128 bit value as octal -> 43 digits - char storage_[128 / 3 + 1]; + private: + const char *start_; + size_t size_; + // Max size: 128 bit value as octal -> 43 digits, plus sign char + char storage_[128 / 3 + 1 + 1]; }; // Note: 'o' conversions do not have a base indicator, it's just that // the '#' flag is specified to modify the precision for 'o' conversions. -string_view BaseIndicator(const ConvertedIntInfo &info, - const ConversionSpec conv) { - bool alt = conv.flags().alt; - int radix = FormatConversionCharRadix(conv.conv()); - if (conv.conv() == ConversionChar::p) alt = true; // always show 0x for %p. +string_view BaseIndicator(const IntDigits &as_digits, + const FormatConversionSpecImpl conv) { + // always show 0x for %p. + bool alt = conv.has_alt_flag() || + conv.conversion_char() == FormatConversionCharInternal::p; + bool hex = (conv.conversion_char() == FormatConversionCharInternal::x || + conv.conversion_char() == FormatConversionCharInternal::X || + conv.conversion_char() == FormatConversionCharInternal::p); // From the POSIX description of '#' flag: // "For x or X conversion specifiers, a non-zero result shall have // 0x (or 0X) prefixed to it." - if (alt && radix == 16 && !info.digits().empty()) { - if (FormatConversionCharIsUpper(conv.conv())) return "0X"; - return "0x"; + if (alt && hex && !as_digits.without_neg_or_zero().empty()) { + return conv.conversion_char() == FormatConversionCharInternal::X ? "0X" + : "0x"; } return {}; } -string_view SignColumn(bool neg, const ConversionSpec conv) { - if (FormatConversionCharIsSigned(conv.conv())) { +string_view SignColumn(bool neg, const FormatConversionSpecImpl conv) { + if (conv.conversion_char() == FormatConversionCharInternal::d || + conv.conversion_char() == FormatConversionCharInternal::i) { if (neg) return "-"; - if (conv.flags().show_pos) return "+"; - if (conv.flags().sign_col) return " "; + if (conv.has_show_pos_flag()) return "+"; + if (conv.has_sign_col_flag()) return " "; } return {}; } -bool ConvertCharImpl(unsigned char v, const ConversionSpec conv, +bool ConvertCharImpl(unsigned char v, const FormatConversionSpecImpl conv, FormatSinkImpl *sink) { size_t fill = 0; if (conv.width() >= 0) fill = conv.width(); ReducePadding(1, &fill); - if (!conv.flags().left) sink->Append(fill, ' '); + if (!conv.has_left_flag()) sink->Append(fill, ' '); sink->Append(1, v); - if (conv.flags().left) sink->Append(fill, ' '); + if (conv.has_left_flag()) sink->Append(fill, ' '); return true; } -bool ConvertIntImplInner(const ConvertedIntInfo &info, - const ConversionSpec conv, FormatSinkImpl *sink) { +bool ConvertIntImplInnerSlow(const IntDigits &as_digits, + const FormatConversionSpecImpl conv, + FormatSinkImpl *sink) { // Print as a sequence of Substrings: // [left_spaces][sign][base_indicator][zeroes][formatted][right_spaces] size_t fill = 0; if (conv.width() >= 0) fill = conv.width(); - string_view formatted = info.digits(); + string_view formatted = as_digits.without_neg_or_zero(); ReducePadding(formatted, &fill); - string_view sign = SignColumn(info.is_neg(), conv); + string_view sign = SignColumn(as_digits.is_negative(), conv); ReducePadding(sign, &fill); - string_view base_indicator = BaseIndicator(info, conv); + string_view base_indicator = BaseIndicator(as_digits, conv); ReducePadding(base_indicator, &fill); int precision = conv.precision(); @@ -174,7 +227,8 @@ bool ConvertIntImplInner(const ConvertedIntInfo &info, if (!precision_specified) precision = 1; - if (conv.flags().alt && conv.conv() == ConversionChar::o) { + if (conv.has_alt_flag() && + conv.conversion_char() == FormatConversionCharInternal::o) { // From POSIX description of the '#' (alt) flag: // "For o conversion, it increases the precision (if necessary) to // force the first digit of the result to be zero." @@ -187,13 +241,13 @@ bool ConvertIntImplInner(const ConvertedIntInfo &info, size_t num_zeroes = Excess(formatted.size(), precision); ReducePadding(num_zeroes, &fill); - size_t num_left_spaces = !conv.flags().left ? fill : 0; - size_t num_right_spaces = conv.flags().left ? fill : 0; + size_t num_left_spaces = !conv.has_left_flag() ? fill : 0; + size_t num_right_spaces = conv.has_left_flag() ? fill : 0; // From POSIX description of the '0' (zero) flag: // "For d, i, o, u, x, and X conversion specifiers, if a precision // is specified, the '0' flag is ignored." - if (!precision_specified && conv.flags().zero) { + if (!precision_specified && conv.has_zero_flag()) { num_zeroes += num_left_spaces; num_left_spaces = 0; } @@ -208,71 +262,97 @@ bool ConvertIntImplInner(const ConvertedIntInfo &info, } template <typename T> -bool ConvertIntImplInner(T v, const ConversionSpec conv, FormatSinkImpl *sink) { - ConvertedIntInfo info(v, conv.conv()); - if (conv.flags().basic && (conv.conv() != ConversionChar::p)) { - if (info.is_neg()) sink->Append(1, '-'); - if (info.digits().empty()) { - sink->Append(1, '0'); - } else { - sink->Append(info.digits()); - } - return true; +bool ConvertIntArg(T v, const FormatConversionSpecImpl conv, + FormatSinkImpl *sink) { + using U = typename MakeUnsigned<T>::type; + IntDigits as_digits; + + // This odd casting is due to a bug in -Wswitch behavior in gcc49 which causes + // it to complain about a switch/case type mismatch, even though both are + // FormatConverionChar. Likely this is because at this point + // FormatConversionChar is declared, but not defined. + switch (static_cast<uint8_t>(conv.conversion_char())) { + case static_cast<uint8_t>(FormatConversionCharInternal::c): + return ConvertCharImpl(static_cast<unsigned char>(v), conv, sink); + + case static_cast<uint8_t>(FormatConversionCharInternal::o): + as_digits.PrintAsOct(static_cast<U>(v)); + break; + + case static_cast<uint8_t>(FormatConversionCharInternal::x): + as_digits.PrintAsHexLower(static_cast<U>(v)); + break; + case static_cast<uint8_t>(FormatConversionCharInternal::X): + as_digits.PrintAsHexUpper(static_cast<U>(v)); + break; + + case static_cast<uint8_t>(FormatConversionCharInternal::u): + as_digits.PrintAsDec(static_cast<U>(v)); + break; + + case static_cast<uint8_t>(FormatConversionCharInternal::d): + case static_cast<uint8_t>(FormatConversionCharInternal::i): + as_digits.PrintAsDec(v); + break; + + case static_cast<uint8_t>(FormatConversionCharInternal::a): + case static_cast<uint8_t>(FormatConversionCharInternal::e): + case static_cast<uint8_t>(FormatConversionCharInternal::f): + case static_cast<uint8_t>(FormatConversionCharInternal::g): + case static_cast<uint8_t>(FormatConversionCharInternal::A): + case static_cast<uint8_t>(FormatConversionCharInternal::E): + case static_cast<uint8_t>(FormatConversionCharInternal::F): + case static_cast<uint8_t>(FormatConversionCharInternal::G): + return ConvertFloatImpl(static_cast<double>(v), conv, sink); + + default: + ABSL_INTERNAL_ASSUME(false); } - return ConvertIntImplInner(info, conv, sink); -} -template <typename T> -bool ConvertIntArg(T v, const ConversionSpec conv, FormatSinkImpl *sink) { - if (FormatConversionCharIsFloat(conv.conv())) { - return FormatConvertImpl(static_cast<double>(v), conv, sink).value; - } - if (conv.conv() == ConversionChar::c) - return ConvertCharImpl(static_cast<unsigned char>(v), conv, sink); - if (!FormatConversionCharIsIntegral(conv.conv())) return false; - if (!FormatConversionCharIsSigned(conv.conv()) && IsSigned<T>::value) { - using U = typename MakeUnsigned<T>::type; - return FormatConvertImpl(static_cast<U>(v), conv, sink).value; + if (conv.is_basic()) { + sink->Append(as_digits.with_neg_and_zero()); + return true; } - return ConvertIntImplInner(v, conv, sink); + return ConvertIntImplInnerSlow(as_digits, conv, sink); } template <typename T> -bool ConvertFloatArg(T v, const ConversionSpec conv, FormatSinkImpl *sink) { - return FormatConversionCharIsFloat(conv.conv()) && +bool ConvertFloatArg(T v, const FormatConversionSpecImpl conv, + FormatSinkImpl *sink) { + return FormatConversionCharIsFloat(conv.conversion_char()) && ConvertFloatImpl(v, conv, sink); } -inline bool ConvertStringArg(string_view v, const ConversionSpec conv, +inline bool ConvertStringArg(string_view v, const FormatConversionSpecImpl conv, FormatSinkImpl *sink) { - if (conv.conv() != ConversionChar::s) return false; - if (conv.flags().basic) { + if (conv.is_basic()) { sink->Append(v); return true; } return sink->PutPaddedString(v, conv.width(), conv.precision(), - conv.flags().left); + conv.has_left_flag()); } } // namespace // ==================== Strings ==================== -ConvertResult<Conv::s> FormatConvertImpl(const std::string &v, - const ConversionSpec conv, - FormatSinkImpl *sink) { +StringConvertResult FormatConvertImpl(const std::string &v, + const FormatConversionSpecImpl conv, + FormatSinkImpl *sink) { return {ConvertStringArg(v, conv, sink)}; } -ConvertResult<Conv::s> FormatConvertImpl(string_view v, - const ConversionSpec conv, - FormatSinkImpl *sink) { +StringConvertResult FormatConvertImpl(string_view v, + const FormatConversionSpecImpl conv, + FormatSinkImpl *sink) { return {ConvertStringArg(v, conv, sink)}; } -ConvertResult<Conv::s | Conv::p> FormatConvertImpl(const char *v, - const ConversionSpec conv, - FormatSinkImpl *sink) { - if (conv.conv() == ConversionChar::p) +ArgConvertResult<FormatConversionCharSetUnion( + FormatConversionCharSetInternal::s, FormatConversionCharSetInternal::p)> +FormatConvertImpl(const char *v, const FormatConversionSpecImpl conv, + FormatSinkImpl *sink) { + if (conv.conversion_char() == FormatConversionCharInternal::p) return {FormatConvertImpl(VoidPtr(v), conv, sink).value}; size_t len; if (v == nullptr) { @@ -287,93 +367,99 @@ ConvertResult<Conv::s | Conv::p> FormatConvertImpl(const char *v, } // ==================== Raw pointers ==================== -ConvertResult<Conv::p> FormatConvertImpl(VoidPtr v, const ConversionSpec conv, - FormatSinkImpl *sink) { - if (conv.conv() != ConversionChar::p) return {false}; +ArgConvertResult<FormatConversionCharSetInternal::p> FormatConvertImpl( + VoidPtr v, const FormatConversionSpecImpl conv, FormatSinkImpl *sink) { if (!v.value) { sink->Append("(nil)"); return {true}; } - return {ConvertIntImplInner(v.value, conv, sink)}; + IntDigits as_digits; + as_digits.PrintAsHexLower(v.value); + return {ConvertIntImplInnerSlow(as_digits, conv, sink)}; } // ==================== Floats ==================== -FloatingConvertResult FormatConvertImpl(float v, const ConversionSpec conv, +FloatingConvertResult FormatConvertImpl(float v, + const FormatConversionSpecImpl conv, FormatSinkImpl *sink) { return {ConvertFloatArg(v, conv, sink)}; } -FloatingConvertResult FormatConvertImpl(double v, const ConversionSpec conv, +FloatingConvertResult FormatConvertImpl(double v, + const FormatConversionSpecImpl conv, FormatSinkImpl *sink) { return {ConvertFloatArg(v, conv, sink)}; } FloatingConvertResult FormatConvertImpl(long double v, - const ConversionSpec conv, + const FormatConversionSpecImpl conv, FormatSinkImpl *sink) { return {ConvertFloatArg(v, conv, sink)}; } // ==================== Chars ==================== -IntegralConvertResult FormatConvertImpl(char v, const ConversionSpec conv, +IntegralConvertResult FormatConvertImpl(char v, + const FormatConversionSpecImpl conv, FormatSinkImpl *sink) { return {ConvertIntArg(v, conv, sink)}; } IntegralConvertResult FormatConvertImpl(signed char v, - const ConversionSpec conv, + const FormatConversionSpecImpl conv, FormatSinkImpl *sink) { return {ConvertIntArg(v, conv, sink)}; } IntegralConvertResult FormatConvertImpl(unsigned char v, - const ConversionSpec conv, + const FormatConversionSpecImpl conv, FormatSinkImpl *sink) { return {ConvertIntArg(v, conv, sink)}; } // ==================== Ints ==================== IntegralConvertResult FormatConvertImpl(short v, // NOLINT - const ConversionSpec conv, + const FormatConversionSpecImpl conv, FormatSinkImpl *sink) { return {ConvertIntArg(v, conv, sink)}; } IntegralConvertResult FormatConvertImpl(unsigned short v, // NOLINT - const ConversionSpec conv, + const FormatConversionSpecImpl conv, FormatSinkImpl *sink) { return {ConvertIntArg(v, conv, sink)}; } -IntegralConvertResult FormatConvertImpl(int v, const ConversionSpec conv, +IntegralConvertResult FormatConvertImpl(int v, + const FormatConversionSpecImpl conv, FormatSinkImpl *sink) { return {ConvertIntArg(v, conv, sink)}; } -IntegralConvertResult FormatConvertImpl(unsigned v, const ConversionSpec conv, +IntegralConvertResult FormatConvertImpl(unsigned v, + const FormatConversionSpecImpl conv, FormatSinkImpl *sink) { return {ConvertIntArg(v, conv, sink)}; } IntegralConvertResult FormatConvertImpl(long v, // NOLINT - const ConversionSpec conv, + const FormatConversionSpecImpl conv, FormatSinkImpl *sink) { return {ConvertIntArg(v, conv, sink)}; } IntegralConvertResult FormatConvertImpl(unsigned long v, // NOLINT - const ConversionSpec conv, + const FormatConversionSpecImpl conv, FormatSinkImpl *sink) { return {ConvertIntArg(v, conv, sink)}; } IntegralConvertResult FormatConvertImpl(long long v, // NOLINT - const ConversionSpec conv, + const FormatConversionSpecImpl conv, FormatSinkImpl *sink) { return {ConvertIntArg(v, conv, sink)}; } IntegralConvertResult FormatConvertImpl(unsigned long long v, // NOLINT - const ConversionSpec conv, + const FormatConversionSpecImpl conv, FormatSinkImpl *sink) { return {ConvertIntArg(v, conv, sink)}; } IntegralConvertResult FormatConvertImpl(absl::int128 v, - const ConversionSpec conv, + const FormatConversionSpecImpl conv, FormatSinkImpl *sink) { return {ConvertIntArg(v, conv, sink)}; } IntegralConvertResult FormatConvertImpl(absl::uint128 v, - const ConversionSpec conv, + const FormatConversionSpecImpl conv, FormatSinkImpl *sink) { return {ConvertIntArg(v, conv, sink)}; } diff --git a/absl/strings/internal/str_format/arg.h b/absl/strings/internal/str_format/arg.h index 7a937563..3dbc1526 100644 --- a/absl/strings/internal/str_format/arg.h +++ b/absl/strings/internal/str_format/arg.h @@ -25,16 +25,37 @@ class Cord; class FormatCountCapture; class FormatSink; +template <absl::FormatConversionCharSet C> +struct FormatConvertResult; +class FormatConversionSpec; + namespace str_format_internal { template <typename T, typename = void> struct HasUserDefinedConvert : std::false_type {}; template <typename T> -struct HasUserDefinedConvert< - T, void_t<decltype(AbslFormatConvert( - std::declval<const T&>(), std::declval<ConversionSpec>(), - std::declval<FormatSink*>()))>> : std::true_type {}; +struct HasUserDefinedConvert<T, void_t<decltype(AbslFormatConvert( + std::declval<const T&>(), + std::declval<const FormatConversionSpec&>(), + std::declval<FormatSink*>()))>> + : std::true_type {}; + +void AbslFormatConvert(); // Stops the lexical name lookup +template <typename T> +auto FormatConvertImpl(const T& v, FormatConversionSpecImpl conv, + FormatSinkImpl* sink) + -> decltype(AbslFormatConvert(v, + std::declval<const FormatConversionSpec&>(), + std::declval<FormatSink*>())) { + using FormatConversionSpecT = + absl::enable_if_t<sizeof(const T& (*)()) != 0, FormatConversionSpec>; + using FormatSinkT = + absl::enable_if_t<sizeof(const T& (*)()) != 0, FormatSink>; + auto fcs = conv.Wrap<FormatConversionSpecT>(); + auto fs = sink->Wrap<FormatSinkT>(); + return AbslFormatConvert(v, fcs, &fs); +} template <typename T> class StreamedWrapper; @@ -43,6 +64,13 @@ class StreamedWrapper; // then convert it, appending to `sink` and return `true`. // Otherwise fail and return `false`. +// AbslFormatConvert(v, conv, sink) is intended to be found by ADL on 'v' +// as an extension mechanism. These FormatConvertImpl functions are the default +// implementations. +// The ADL search is augmented via the 'Sink*' parameter, which also +// serves as a disambiguator to reject possible unintended 'AbslFormatConvert' +// functions in the namespaces associated with 'v'. + // Raw pointers. struct VoidPtr { VoidPtr() = default; @@ -52,27 +80,45 @@ struct VoidPtr { : value(ptr ? reinterpret_cast<uintptr_t>(ptr) : 0) {} uintptr_t value; }; -ConvertResult<Conv::p> FormatConvertImpl(VoidPtr v, ConversionSpec conv, - FormatSinkImpl* sink); + +template <FormatConversionCharSet C> +struct ArgConvertResult { + bool value; +}; + +template <FormatConversionCharSet C> +constexpr FormatConversionCharSet ExtractCharSet(FormatConvertResult<C>) { + return C; +} + +template <FormatConversionCharSet C> +constexpr FormatConversionCharSet ExtractCharSet(ArgConvertResult<C>) { + return C; +} + +using StringConvertResult = + ArgConvertResult<FormatConversionCharSetInternal::s>; +ArgConvertResult<FormatConversionCharSetInternal::p> FormatConvertImpl( + VoidPtr v, FormatConversionSpecImpl conv, FormatSinkImpl* sink); // Strings. -ConvertResult<Conv::s> FormatConvertImpl(const std::string& v, - ConversionSpec conv, - FormatSinkImpl* sink); -ConvertResult<Conv::s> FormatConvertImpl(string_view v, ConversionSpec conv, - FormatSinkImpl* sink); -ConvertResult<Conv::s | Conv::p> FormatConvertImpl(const char* v, - ConversionSpec conv, - FormatSinkImpl* sink); -template <class AbslCord, - typename std::enable_if< - std::is_same<AbslCord, absl::Cord>::value>::type* = nullptr> -ConvertResult<Conv::s> FormatConvertImpl(const AbslCord& value, - ConversionSpec conv, - FormatSinkImpl* sink) { - if (conv.conv() != ConversionChar::s) return {false}; - - bool is_left = conv.flags().left; +StringConvertResult FormatConvertImpl(const std::string& v, + FormatConversionSpecImpl conv, + FormatSinkImpl* sink); +StringConvertResult FormatConvertImpl(string_view v, + FormatConversionSpecImpl conv, + FormatSinkImpl* sink); +ArgConvertResult<FormatConversionCharSetUnion( + FormatConversionCharSetInternal::s, FormatConversionCharSetInternal::p)> +FormatConvertImpl(const char* v, const FormatConversionSpecImpl conv, + FormatSinkImpl* sink); + +template <class AbslCord, typename std::enable_if<std::is_same< + AbslCord, absl::Cord>::value>::type* = nullptr> +StringConvertResult FormatConvertImpl(const AbslCord& value, + FormatConversionSpecImpl conv, + FormatSinkImpl* sink) { + bool is_left = conv.has_left_flag(); size_t space_remaining = 0; int width = conv.width(); @@ -105,55 +151,63 @@ ConvertResult<Conv::s> FormatConvertImpl(const AbslCord& value, return {true}; } -using IntegralConvertResult = - ConvertResult<Conv::c | Conv::numeric | Conv::star>; -using FloatingConvertResult = ConvertResult<Conv::floating>; +using IntegralConvertResult = ArgConvertResult<FormatConversionCharSetUnion( + FormatConversionCharSetInternal::c, + FormatConversionCharSetInternal::kNumeric, + FormatConversionCharSetInternal::kStar)>; +using FloatingConvertResult = + ArgConvertResult<FormatConversionCharSetInternal::kFloating>; // Floats. -FloatingConvertResult FormatConvertImpl(float v, ConversionSpec conv, +FloatingConvertResult FormatConvertImpl(float v, FormatConversionSpecImpl conv, FormatSinkImpl* sink); -FloatingConvertResult FormatConvertImpl(double v, ConversionSpec conv, +FloatingConvertResult FormatConvertImpl(double v, FormatConversionSpecImpl conv, FormatSinkImpl* sink); -FloatingConvertResult FormatConvertImpl(long double v, ConversionSpec conv, +FloatingConvertResult FormatConvertImpl(long double v, + FormatConversionSpecImpl conv, FormatSinkImpl* sink); // Chars. -IntegralConvertResult FormatConvertImpl(char v, ConversionSpec conv, +IntegralConvertResult FormatConvertImpl(char v, FormatConversionSpecImpl conv, FormatSinkImpl* sink); -IntegralConvertResult FormatConvertImpl(signed char v, ConversionSpec conv, +IntegralConvertResult FormatConvertImpl(signed char v, + FormatConversionSpecImpl conv, FormatSinkImpl* sink); -IntegralConvertResult FormatConvertImpl(unsigned char v, ConversionSpec conv, +IntegralConvertResult FormatConvertImpl(unsigned char v, + FormatConversionSpecImpl conv, FormatSinkImpl* sink); // Ints. IntegralConvertResult FormatConvertImpl(short v, // NOLINT - ConversionSpec conv, + FormatConversionSpecImpl conv, FormatSinkImpl* sink); IntegralConvertResult FormatConvertImpl(unsigned short v, // NOLINT - ConversionSpec conv, + FormatConversionSpecImpl conv, FormatSinkImpl* sink); -IntegralConvertResult FormatConvertImpl(int v, ConversionSpec conv, +IntegralConvertResult FormatConvertImpl(int v, FormatConversionSpecImpl conv, FormatSinkImpl* sink); -IntegralConvertResult FormatConvertImpl(unsigned v, ConversionSpec conv, +IntegralConvertResult FormatConvertImpl(unsigned v, + FormatConversionSpecImpl conv, FormatSinkImpl* sink); IntegralConvertResult FormatConvertImpl(long v, // NOLINT - ConversionSpec conv, + FormatConversionSpecImpl conv, FormatSinkImpl* sink); IntegralConvertResult FormatConvertImpl(unsigned long v, // NOLINT - ConversionSpec conv, + FormatConversionSpecImpl conv, FormatSinkImpl* sink); IntegralConvertResult FormatConvertImpl(long long v, // NOLINT - ConversionSpec conv, + FormatConversionSpecImpl conv, FormatSinkImpl* sink); IntegralConvertResult FormatConvertImpl(unsigned long long v, // NOLINT - ConversionSpec conv, + FormatConversionSpecImpl conv, FormatSinkImpl* sink); -IntegralConvertResult FormatConvertImpl(int128 v, ConversionSpec conv, +IntegralConvertResult FormatConvertImpl(int128 v, FormatConversionSpecImpl conv, FormatSinkImpl* sink); -IntegralConvertResult FormatConvertImpl(uint128 v, ConversionSpec conv, +IntegralConvertResult FormatConvertImpl(uint128 v, + FormatConversionSpecImpl conv, FormatSinkImpl* sink); template <typename T, enable_if_t<std::is_same<T, bool>::value, int> = 0> -IntegralConvertResult FormatConvertImpl(T v, ConversionSpec conv, +IntegralConvertResult FormatConvertImpl(T v, FormatConversionSpecImpl conv, FormatSinkImpl* sink) { return FormatConvertImpl(static_cast<int>(v), conv, sink); } @@ -164,12 +218,12 @@ template <typename T> typename std::enable_if<std::is_enum<T>::value && !HasUserDefinedConvert<T>::value, IntegralConvertResult>::type -FormatConvertImpl(T v, ConversionSpec conv, FormatSinkImpl* sink); +FormatConvertImpl(T v, FormatConversionSpecImpl conv, FormatSinkImpl* sink); template <typename T> -ConvertResult<Conv::s> FormatConvertImpl(const StreamedWrapper<T>& v, - ConversionSpec conv, - FormatSinkImpl* out) { +StringConvertResult FormatConvertImpl(const StreamedWrapper<T>& v, + FormatConversionSpecImpl conv, + FormatSinkImpl* out) { std::ostringstream oss; oss << v.v_; if (!oss) return {false}; @@ -180,21 +234,24 @@ ConvertResult<Conv::s> FormatConvertImpl(const StreamedWrapper<T>& v, // until after FormatCountCapture is fully defined. struct FormatCountCaptureHelper { template <class T = int> - static ConvertResult<Conv::n> ConvertHelper(const FormatCountCapture& v, - ConversionSpec conv, - FormatSinkImpl* sink) { + static ArgConvertResult<FormatConversionCharSetInternal::n> ConvertHelper( + const FormatCountCapture& v, FormatConversionSpecImpl conv, + FormatSinkImpl* sink) { const absl::enable_if_t<sizeof(T) != 0, FormatCountCapture>& v2 = v; - if (conv.conv() != str_format_internal::ConversionChar::n) return {false}; + if (conv.conversion_char() != + str_format_internal::FormatConversionCharInternal::n) { + return {false}; + } *v2.p_ = static_cast<int>(sink->size()); return {true}; } }; template <class T = int> -ConvertResult<Conv::n> FormatConvertImpl(const FormatCountCapture& v, - ConversionSpec conv, - FormatSinkImpl* sink) { +ArgConvertResult<FormatConversionCharSetInternal::n> FormatConvertImpl( + const FormatCountCapture& v, FormatConversionSpecImpl conv, + FormatSinkImpl* sink) { return FormatCountCaptureHelper::ConvertHelper(v, conv, sink); } @@ -203,13 +260,13 @@ ConvertResult<Conv::n> FormatConvertImpl(const FormatCountCapture& v, struct FormatArgImplFriend { template <typename Arg> static bool ToInt(Arg arg, int* out) { - // A value initialized ConversionSpec has a `none` conv, which tells the - // dispatcher to run the `int` conversion. + // A value initialized FormatConversionSpecImpl has a `none` conv, which + // tells the dispatcher to run the `int` conversion. return arg.dispatcher_(arg.data_, {}, out); } template <typename Arg> - static bool Convert(Arg arg, str_format_internal::ConversionSpec conv, + static bool Convert(Arg arg, FormatConversionSpecImpl conv, FormatSinkImpl* out) { return arg.dispatcher_(arg.data_, conv, out); } @@ -220,6 +277,15 @@ struct FormatArgImplFriend { } }; +template <typename Arg> +constexpr FormatConversionCharSet ArgumentToConv() { + return absl::str_format_internal::ExtractCharSet( + decltype(str_format_internal::FormatConvertImpl( + std::declval<const Arg&>(), + std::declval<const FormatConversionSpecImpl&>(), + std::declval<FormatSinkImpl*>())){}); +} + // A type-erased handle to a format argument. class FormatArgImpl { private: @@ -233,7 +299,7 @@ class FormatArgImpl { char buf[kInlinedSpace]; }; - using Dispatcher = bool (*)(Data, ConversionSpec, void* out); + using Dispatcher = bool (*)(Data, FormatConversionSpecImpl, void* out); template <typename T> struct store_by_value @@ -375,15 +441,20 @@ class FormatArgImpl { } template <typename T> - static bool Dispatch(Data arg, ConversionSpec spec, void* out) { + static bool Dispatch(Data arg, FormatConversionSpecImpl spec, void* out) { // A `none` conv indicates that we want the `int` conversion. - if (ABSL_PREDICT_FALSE(spec.conv() == ConversionChar::none)) { + if (ABSL_PREDICT_FALSE(spec.conversion_char() == + FormatConversionCharInternal::kNone)) { return ToInt<T>(arg, static_cast<int*>(out), std::is_integral<T>(), std::is_enum<T>()); } - + if (ABSL_PREDICT_FALSE(!Contains(ArgumentToConv<T>(), + spec.conversion_char()))) { + return false; + } return str_format_internal::FormatConvertImpl( - Manager<T>::Value(arg), spec, static_cast<FormatSinkImpl*>(out)) + Manager<T>::Value(arg), spec, + static_cast<FormatSinkImpl*>(out)) .value; } @@ -391,8 +462,9 @@ class FormatArgImpl { Dispatcher dispatcher_; }; -#define ABSL_INTERNAL_FORMAT_DISPATCH_INSTANTIATE_(T, E) \ - E template bool FormatArgImpl::Dispatch<T>(Data, ConversionSpec, void*) +#define ABSL_INTERNAL_FORMAT_DISPATCH_INSTANTIATE_(T, E) \ + E template bool FormatArgImpl::Dispatch<T>(Data, FormatConversionSpecImpl, \ + void*) #define ABSL_INTERNAL_FORMAT_DISPATCH_OVERLOADS_EXPAND_(...) \ ABSL_INTERNAL_FORMAT_DISPATCH_INSTANTIATE_(str_format_internal::VoidPtr, \ diff --git a/absl/strings/internal/str_format/arg_test.cc b/absl/strings/internal/str_format/arg_test.cc index 8d30d8b8..f53fd6bd 100644 --- a/absl/strings/internal/str_format/arg_test.cc +++ b/absl/strings/internal/str_format/arg_test.cc @@ -23,8 +23,17 @@ class FormatArgImplTest : public ::testing::Test { enum Color { kRed, kGreen, kBlue }; static const char *hi() { return "hi"; } + + struct X {}; + + X x_; }; +inline FormatConvertResult<FormatConversionCharSet{}> AbslFormatConvert( + const FormatArgImplTest::X &, const FormatConversionSpec &, FormatSink *) { + return {false}; +} + TEST_F(FormatArgImplTest, ToInt) { int out = 0; EXPECT_TRUE(FormatArgImplFriend::ToInt(FormatArgImpl(1), &out)); @@ -59,6 +68,7 @@ TEST_F(FormatArgImplTest, ToInt) { FormatArgImpl(static_cast<int *>(nullptr)), &out)); EXPECT_FALSE(FormatArgImplFriend::ToInt(FormatArgImpl(hi()), &out)); EXPECT_FALSE(FormatArgImplFriend::ToInt(FormatArgImpl("hi"), &out)); + EXPECT_FALSE(FormatArgImplFriend::ToInt(FormatArgImpl(x_), &out)); EXPECT_TRUE(FormatArgImplFriend::ToInt(FormatArgImpl(kBlue), &out)); EXPECT_EQ(2, out); } @@ -95,8 +105,9 @@ TEST_F(FormatArgImplTest, OtherPtrDecayToVoidPtr) { TEST_F(FormatArgImplTest, WorksWithCharArraysOfUnknownSize) { std::string s; FormatSinkImpl sink(&s); - ConversionSpec conv; - FormatConversionSpecImplFriend::SetConversionChar(ConversionChar::s, &conv); + FormatConversionSpecImpl conv; + FormatConversionSpecImplFriend::SetConversionChar( + FormatConversionCharInternal::s, &conv); FormatConversionSpecImplFriend::SetFlags(Flags(), &conv); FormatConversionSpecImplFriend::SetWidth(-1, &conv); FormatConversionSpecImplFriend::SetPrecision(-1, &conv); diff --git a/absl/strings/internal/str_format/bind.cc b/absl/strings/internal/str_format/bind.cc index 27522fdb..6980ed1d 100644 --- a/absl/strings/internal/str_format/bind.cc +++ b/absl/strings/internal/str_format/bind.cc @@ -147,7 +147,7 @@ class SummarizingConverter { << FormatConversionSpecImplFriend::FlagsToString(bound); if (bound.width() >= 0) ss << bound.width(); if (bound.precision() >= 0) ss << "." << bound.precision(); - ss << bound.conv() << "}"; + ss << bound.conversion_char() << "}"; Append(ss.str()); return true; } diff --git a/absl/strings/internal/str_format/bind.h b/absl/strings/internal/str_format/bind.h index cf41b197..585246e7 100644 --- a/absl/strings/internal/str_format/bind.h +++ b/absl/strings/internal/str_format/bind.h @@ -19,7 +19,7 @@ class UntypedFormatSpec; namespace str_format_internal { -class BoundConversion : public ConversionSpec { +class BoundConversion : public FormatConversionSpecImpl { public: const FormatArgImpl* arg() const { return arg_; } void set_arg(const FormatArgImpl* a) { arg_ = a; } @@ -60,7 +60,7 @@ class UntypedFormatSpecImpl { size_t size_; }; -template <typename T, typename...> +template <typename T, FormatConversionCharSet...> struct MakeDependent { using type = T; }; @@ -68,7 +68,7 @@ struct MakeDependent { // Implicitly convertible from `const char*`, `string_view`, and the // `ExtendedParsedFormat` type. This abstraction allows all format functions to // operate on any without providing too many overloads. -template <typename... Args> +template <FormatConversionCharSet... Args> class FormatSpecTemplate : public MakeDependent<UntypedFormatSpec, Args...>::type { using Base = typename MakeDependent<UntypedFormatSpec, Args...>::type; @@ -76,11 +76,11 @@ class FormatSpecTemplate public: #ifdef ABSL_INTERNAL_ENABLE_FORMAT_CHECKER - // Honeypot overload for when the std::string is not constexpr. + // Honeypot overload for when the string is not constexpr. // We use the 'unavailable' attribute to give a better compiler error than // just 'method is deleted'. FormatSpecTemplate(...) // NOLINT - __attribute__((unavailable("Format std::string is not constexpr."))); + __attribute__((unavailable("Format string is not constexpr."))); // Honeypot overload for when the format is constexpr and invalid. // We use the 'unavailable' attribute to give a better compiler error than @@ -105,13 +105,11 @@ class FormatSpecTemplate // Good format overload. FormatSpecTemplate(const char* s) // NOLINT - __attribute__((enable_if(ValidFormatImpl<ArgumentToConv<Args>()...>(s), - "bad format trap"))) + __attribute__((enable_if(ValidFormatImpl<Args...>(s), "bad format trap"))) : Base(s) {} FormatSpecTemplate(string_view s) // NOLINT - __attribute__((enable_if(ValidFormatImpl<ArgumentToConv<Args>()...>(s), - "bad format trap"))) + __attribute__((enable_if(ValidFormatImpl<Args...>(s), "bad format trap"))) : Base(s) {} #else // ABSL_INTERNAL_ENABLE_FORMAT_CHECKER @@ -121,19 +119,14 @@ class FormatSpecTemplate #endif // ABSL_INTERNAL_ENABLE_FORMAT_CHECKER - template <Conv... C, typename = typename std::enable_if< - AllOf(sizeof...(C) == sizeof...(Args), - Contains(ArgumentToConv<Args>(), - C)...)>::type> + template <FormatConversionCharSet... C, + typename = typename std::enable_if< + AllOf(sizeof...(C) == sizeof...(Args), Contains(Args, + C)...)>::type> FormatSpecTemplate(const ExtendedParsedFormat<C...>& pc) // NOLINT : Base(&pc) {} }; -template <typename... Args> -struct FormatSpecDeductionBarrier { - using type = FormatSpecTemplate<Args...>; -}; - class Streamable { public: Streamable(const UntypedFormatSpecImpl& format, @@ -196,9 +189,9 @@ class StreamedWrapper { private: template <typename S> - friend ConvertResult<Conv::s> FormatConvertImpl(const StreamedWrapper<S>& v, - ConversionSpec conv, - FormatSinkImpl* out); + friend ArgConvertResult<FormatConversionCharSetInternal::s> FormatConvertImpl( + const StreamedWrapper<S>& v, FormatConversionSpecImpl conv, + FormatSinkImpl* out); const T& v_; }; diff --git a/absl/strings/internal/str_format/checker.h b/absl/strings/internal/str_format/checker.h index 8993a79b..424c51f7 100644 --- a/absl/strings/internal/str_format/checker.h +++ b/absl/strings/internal/str_format/checker.h @@ -24,13 +24,6 @@ constexpr bool AllOf(bool b, T... t) { return b && AllOf(t...); } -template <typename Arg> -constexpr Conv ArgumentToConv() { - return decltype(str_format_internal::FormatConvertImpl( - std::declval<const Arg&>(), std::declval<const ConversionSpec&>(), - std::declval<FormatSinkImpl*>()))::kConv; -} - #ifdef ABSL_INTERNAL_ENABLE_FORMAT_CHECKER constexpr bool ContainsChar(const char* chars, char c) { @@ -39,14 +32,14 @@ constexpr bool ContainsChar(const char* chars, char c) { // A constexpr compatible list of Convs. struct ConvList { - const Conv* array; + const FormatConversionCharSet* array; int count; // We do the bound check here to avoid having to do it on the callers. - // Returning an empty Conv has the same effect as short circuiting because it - // will never match any conversion. - constexpr Conv operator[](int i) const { - return i < count ? array[i] : Conv{}; + // Returning an empty FormatConversionCharSet has the same effect as + // short circuiting because it will never match any conversion. + constexpr FormatConversionCharSet operator[](int i) const { + return i < count ? array[i] : FormatConversionCharSet{}; } constexpr ConvList without_front() const { @@ -57,7 +50,7 @@ struct ConvList { template <size_t count> struct ConvListT { // Make sure the array has size > 0. - Conv list[count ? count : 1]; + FormatConversionCharSet list[count ? count : 1]; }; constexpr char GetChar(string_view str, size_t index) { @@ -310,7 +303,7 @@ class FormatParser { ConvList args_; }; -template <Conv... C> +template <FormatConversionCharSet... C> constexpr bool ValidFormatImpl(string_view format) { return FormatParser(format, {ConvListT<sizeof...(C)>{{C...}}.list, sizeof...(C)}) diff --git a/absl/strings/internal/str_format/checker_test.cc b/absl/strings/internal/str_format/checker_test.cc index ea2a7681..a76d70b0 100644 --- a/absl/strings/internal/str_format/checker_test.cc +++ b/absl/strings/internal/str_format/checker_test.cc @@ -9,18 +9,22 @@ ABSL_NAMESPACE_BEGIN namespace str_format_internal { namespace { -std::string ConvToString(Conv conv) { +std::string ConvToString(FormatConversionCharSet conv) { std::string out; -#define CONV_SET_CASE(c) \ - if (Contains(conv, Conv::c)) out += #c; +#define CONV_SET_CASE(c) \ + if (Contains(conv, FormatConversionCharSetInternal::c)) { \ + out += #c; \ + } ABSL_INTERNAL_CONVERSION_CHARS_EXPAND_(CONV_SET_CASE, ) #undef CONV_SET_CASE - if (Contains(conv, Conv::star)) out += "*"; + if (Contains(conv, FormatConversionCharSetInternal::kStar)) { + out += "*"; + } return out; } TEST(StrFormatChecker, ArgumentToConv) { - Conv conv = ArgumentToConv<std::string>(); + FormatConversionCharSet conv = ArgumentToConv<std::string>(); EXPECT_EQ(ConvToString(conv), "s"); conv = ArgumentToConv<const char*>(); diff --git a/absl/strings/internal/str_format/convert_test.cc b/absl/strings/internal/str_format/convert_test.cc index cbcd7caf..634ee78b 100644 --- a/absl/strings/internal/str_format/convert_test.cc +++ b/absl/strings/internal/str_format/convert_test.cc @@ -1,20 +1,32 @@ #include <errno.h> #include <stdarg.h> #include <stdio.h> + #include <cctype> #include <cmath> +#include <limits> #include <string> +#include <thread> // NOLINT #include "gmock/gmock.h" #include "gtest/gtest.h" #include "absl/base/internal/raw_logging.h" #include "absl/strings/internal/str_format/bind.h" +#include "absl/strings/match.h" +#include "absl/types/optional.h" namespace absl { ABSL_NAMESPACE_BEGIN namespace str_format_internal { namespace { +struct NativePrintfTraits { + bool hex_float_has_glibc_rounding; + bool hex_float_prefers_denormal_repr; + bool hex_float_uses_minimal_precision_when_not_specified; + bool hex_float_optimizes_leading_digit_bit_count; +}; + template <typename T, size_t N> size_t ArraySize(T (&)[N]) { return N; @@ -57,7 +69,7 @@ std::string Esc(const T &v) { return oss.str(); } -void StrAppend(std::string *dst, const char *format, va_list ap) { +void StrAppendV(std::string *dst, const char *format, va_list ap) { // First try with a small fixed size buffer static const int kSpaceLength = 1024; char space[kSpaceLength]; @@ -98,15 +110,79 @@ void StrAppend(std::string *dst, const char *format, va_list ap) { delete[] buf; } +void StrAppend(std::string *out, const char *format, ...) { + va_list ap; + va_start(ap, format); + StrAppendV(out, format, ap); + va_end(ap); +} + std::string StrPrint(const char *format, ...) { va_list ap; va_start(ap, format); std::string result; - StrAppend(&result, format, ap); + StrAppendV(&result, format, ap); va_end(ap); return result; } +NativePrintfTraits VerifyNativeImplementationImpl() { + NativePrintfTraits result; + + // >>> hex_float_has_glibc_rounding. To have glibc's rounding behavior we need + // to meet three requirements: + // + // - The threshold for rounding up is 8 (for e.g. MSVC uses 9). + // - If the digits lower than than the 8 are non-zero then we round up. + // - If the digits lower than the 8 are all zero then we round toward even. + // + // The numbers below represent all the cases covering {below,at,above} the + // threshold (8) with both {zero,non-zero} lower bits and both {even,odd} + // preceding digits. + const double d0079 = 65657.0; // 0x1.0079p+16 + const double d0179 = 65913.0; // 0x1.0179p+16 + const double d0080 = 65664.0; // 0x1.0080p+16 + const double d0180 = 65920.0; // 0x1.0180p+16 + const double d0081 = 65665.0; // 0x1.0081p+16 + const double d0181 = 65921.0; // 0x1.0181p+16 + result.hex_float_has_glibc_rounding = + StartsWith(StrPrint("%.2a", d0079), "0x1.00") && + StartsWith(StrPrint("%.2a", d0179), "0x1.01") && + StartsWith(StrPrint("%.2a", d0080), "0x1.00") && + StartsWith(StrPrint("%.2a", d0180), "0x1.02") && + StartsWith(StrPrint("%.2a", d0081), "0x1.01") && + StartsWith(StrPrint("%.2a", d0181), "0x1.02"); + + // >>> hex_float_prefers_denormal_repr. Formatting `denormal` on glibc yields + // "0x0.0000000000001p-1022", whereas on std libs that don't use denormal + // representation it would either be 0x1p-1074 or 0x1.0000000000000-1074. + const double denormal = std::numeric_limits<double>::denorm_min(); + result.hex_float_prefers_denormal_repr = + StartsWith(StrPrint("%a", denormal), "0x0.0000000000001"); + + // >>> hex_float_uses_minimal_precision_when_not_specified. Some (non-glibc) + // libs will format the following as "0x1.0079000000000p+16". + result.hex_float_uses_minimal_precision_when_not_specified = + (StrPrint("%a", d0079) == "0x1.0079p+16"); + + // >>> hex_float_optimizes_leading_digit_bit_count. The number 1.5, when + // formatted by glibc should yield "0x1.8p+0" for `double` and "0xcp-3" for + // `long double`, i.e., number of bits in the leading digit is adapted to the + // number of bits in the mantissa. + const double d_15 = 1.5; + const long double ld_15 = 1.5; + result.hex_float_optimizes_leading_digit_bit_count = + StartsWith(StrPrint("%a", d_15), "0x1.8") && + StartsWith(StrPrint("%La", ld_15), "0xc"); + + return result; +} + +const NativePrintfTraits &VerifyNativeImplementation() { + static NativePrintfTraits native_traits = VerifyNativeImplementationImpl(); + return native_traits; +} + class FormatConvertTest : public ::testing::Test { }; template <typename T> @@ -463,6 +539,68 @@ TEST_F(FormatConvertTest, Uint128) { } } +template <typename Floating> +void TestWithMultipleFormatsHelper(const std::vector<Floating> &floats) { + const NativePrintfTraits &native_traits = VerifyNativeImplementation(); + // Reserve the space to ensure we don't allocate memory in the output itself. + std::string str_format_result; + str_format_result.reserve(1 << 20); + std::string string_printf_result; + string_printf_result.reserve(1 << 20); + + const char *const kFormats[] = { + "%", "%.3", "%8.5", "%500", "%.5000", "%.60", "%.30", "%03", + "%+", "% ", "%-10", "%#15.3", "%#.0", "%.0", "%1$*2$", "%1$.*2$"}; + + for (const char *fmt : kFormats) { + for (char f : {'f', 'F', // + 'g', 'G', // + 'a', 'A', // + 'e', 'E'}) { + std::string fmt_str = std::string(fmt) + f; + + if (fmt == absl::string_view("%.5000") && f != 'f' && f != 'F' && + f != 'a' && f != 'A') { + // This particular test takes way too long with snprintf. + // Disable for the case we are not implementing natively. + continue; + } + + if ((f == 'a' || f == 'A') && + !native_traits.hex_float_has_glibc_rounding) { + continue; + } + + for (Floating d : floats) { + if (!native_traits.hex_float_prefers_denormal_repr && + (f == 'a' || f == 'A') && std::fpclassify(d) == FP_SUBNORMAL) { + continue; + } + int i = -10; + FormatArgImpl args[2] = {FormatArgImpl(d), FormatArgImpl(i)}; + UntypedFormatSpecImpl format(fmt_str); + + string_printf_result.clear(); + StrAppend(&string_printf_result, fmt_str.c_str(), d, i); + str_format_result.clear(); + + { + AppendPack(&str_format_result, format, absl::MakeSpan(args)); + } + + if (string_printf_result != str_format_result) { + // We use ASSERT_EQ here because failures are usually correlated and a + // bug would print way too many failed expectations causing the test + // to time out. + ASSERT_EQ(string_printf_result, str_format_result) + << fmt_str << " " << StrPrint("%.18g", d) << " " + << StrPrint("%a", d) << " " << StrPrint("%.50f", d); + } + } + } + } +} + TEST_F(FormatConvertTest, Float) { #ifdef _MSC_VER // MSVC has a different rounding policy than us so we can't test our @@ -470,9 +608,62 @@ TEST_F(FormatConvertTest, Float) { return; #endif // _MSC_VER - const char *const kFormats[] = { - "%", "%.3", "%8.5", "%9", "%.60", "%.30", "%03", "%+", - "% ", "%-10", "%#15.3", "%#.0", "%.0", "%1$*2$", "%1$.*2$"}; + std::vector<float> floats = {0.0f, + -0.0f, + .9999999f, + 9999999.f, + std::numeric_limits<float>::max(), + -std::numeric_limits<float>::max(), + std::numeric_limits<float>::min(), + -std::numeric_limits<float>::min(), + std::numeric_limits<float>::lowest(), + -std::numeric_limits<float>::lowest(), + std::numeric_limits<float>::epsilon(), + std::numeric_limits<float>::epsilon() + 1.0f, + std::numeric_limits<float>::infinity(), + -std::numeric_limits<float>::infinity()}; + + // Some regression tests. + floats.push_back(0.999999989f); + + if (std::numeric_limits<float>::has_denorm != std::denorm_absent) { + floats.push_back(std::numeric_limits<float>::denorm_min()); + floats.push_back(-std::numeric_limits<float>::denorm_min()); + } + + for (float base : + {1.f, 12.f, 123.f, 1234.f, 12345.f, 123456.f, 1234567.f, 12345678.f, + 123456789.f, 1234567890.f, 12345678901.f, 12345678.f, 12345678.f}) { + for (int exp = -123; exp <= 123; ++exp) { + for (int sign : {1, -1}) { + floats.push_back(sign * std::ldexp(base, exp)); + } + } + } + + for (int exp = -300; exp <= 300; ++exp) { + const float all_ones_mantissa = 0xffffff; + floats.push_back(std::ldexp(all_ones_mantissa, exp)); + } + + // Remove duplicates to speed up the logic below. + std::sort(floats.begin(), floats.end()); + floats.erase(std::unique(floats.begin(), floats.end()), floats.end()); + +#ifndef __APPLE__ + // Apple formats NaN differently (+nan) vs. (nan) + floats.push_back(std::nan("")); +#endif + + TestWithMultipleFormatsHelper(floats); +} + +TEST_F(FormatConvertTest, Double) { +#ifdef _MSC_VER + // MSVC has a different rounding policy than us so we can't test our + // implementation against the native one there. + return; +#endif // _MSC_VER std::vector<double> doubles = {0.0, -0.0, @@ -489,11 +680,6 @@ TEST_F(FormatConvertTest, Float) { std::numeric_limits<double>::infinity(), -std::numeric_limits<double>::infinity()}; -#ifndef __APPLE__ - // Apple formats NaN differently (+nan) vs. (nan) - doubles.push_back(std::nan("")); -#endif - // Some regression tests. doubles.push_back(0.99999999999999989); @@ -512,43 +698,375 @@ TEST_F(FormatConvertTest, Float) { } } - for (const char *fmt : kFormats) { - for (char f : {'f', 'F', // - 'g', 'G', // - 'a', 'A', // - 'e', 'E'}) { - std::string fmt_str = std::string(fmt) + f; - for (double d : doubles) { - int i = -10; - FormatArgImpl args[2] = {FormatArgImpl(d), FormatArgImpl(i)}; - UntypedFormatSpecImpl format(fmt_str); - // We use ASSERT_EQ here because failures are usually correlated and a - // bug would print way too many failed expectations causing the test to - // time out. - ASSERT_EQ(StrPrint(fmt_str.c_str(), d, i), - FormatPack(format, absl::MakeSpan(args))) - << fmt_str << " " << StrPrint("%.18g", d) << " " - << StrPrint("%.999f", d); + // Workaround libc bug. + // https://sourceware.org/bugzilla/show_bug.cgi?id=22142 + const bool gcc_bug_22142 = + StrPrint("%f", std::numeric_limits<double>::max()) != + "1797693134862315708145274237317043567980705675258449965989174768031" + "5726078002853876058955863276687817154045895351438246423432132688946" + "4182768467546703537516986049910576551282076245490090389328944075868" + "5084551339423045832369032229481658085593321233482747978262041447231" + "68738177180919299881250404026184124858368.000000"; + + if (!gcc_bug_22142) { + for (int exp = -300; exp <= 300; ++exp) { + const double all_ones_mantissa = 0x1fffffffffffff; + doubles.push_back(std::ldexp(all_ones_mantissa, exp)); + } + } + + if (gcc_bug_22142) { + for (auto &d : doubles) { + using L = std::numeric_limits<double>; + double d2 = std::abs(d); + if (d2 == L::max() || d2 == L::min() || d2 == L::denorm_min()) { + d = 0; } } } + + // Remove duplicates to speed up the logic below. + std::sort(doubles.begin(), doubles.end()); + doubles.erase(std::unique(doubles.begin(), doubles.end()), doubles.end()); + +#ifndef __APPLE__ + // Apple formats NaN differently (+nan) vs. (nan) + doubles.push_back(std::nan("")); +#endif + + TestWithMultipleFormatsHelper(doubles); +} + +TEST_F(FormatConvertTest, DoubleRound) { + std::string s; + const auto format = [&](const char *fmt, double d) -> std::string & { + s.clear(); + FormatArgImpl args[1] = {FormatArgImpl(d)}; + AppendPack(&s, UntypedFormatSpecImpl(fmt), absl::MakeSpan(args)); +#if !defined(_MSC_VER) + // MSVC has a different rounding policy than us so we can't test our + // implementation against the native one there. + EXPECT_EQ(StrPrint(fmt, d), s); +#endif // _MSC_VER + + return s; + }; + // All of these values have to be exactly represented. + // Otherwise we might not be testing what we think we are testing. + + // These values can fit in a 64bit "fast" representation. + const double exact_value = 0.00000000000005684341886080801486968994140625; + assert(exact_value == std::pow(2, -44)); + // Round up at a 5xx. + EXPECT_EQ(format("%.13f", exact_value), "0.0000000000001"); + // Round up at a >5 + EXPECT_EQ(format("%.14f", exact_value), "0.00000000000006"); + // Round down at a <5 + EXPECT_EQ(format("%.16f", exact_value), "0.0000000000000568"); + // Nine handling + EXPECT_EQ(format("%.35f", exact_value), + "0.00000000000005684341886080801486969"); + EXPECT_EQ(format("%.36f", exact_value), + "0.000000000000056843418860808014869690"); + // Round down the last nine. + EXPECT_EQ(format("%.37f", exact_value), + "0.0000000000000568434188608080148696899"); + EXPECT_EQ(format("%.10f", 0.000003814697265625), "0.0000038147"); + // Round up the last nine + EXPECT_EQ(format("%.11f", 0.000003814697265625), "0.00000381470"); + EXPECT_EQ(format("%.12f", 0.000003814697265625), "0.000003814697"); + + // Round to even (down) + EXPECT_EQ(format("%.43f", exact_value), + "0.0000000000000568434188608080148696899414062"); + // Exact + EXPECT_EQ(format("%.44f", exact_value), + "0.00000000000005684341886080801486968994140625"); + // Round to even (up), let make the last digits 75 instead of 25 + EXPECT_EQ(format("%.43f", exact_value + std::pow(2, -43)), + "0.0000000000001705302565824240446090698242188"); + // Exact, just to check. + EXPECT_EQ(format("%.44f", exact_value + std::pow(2, -43)), + "0.00000000000017053025658242404460906982421875"); + + // This value has to be small enough that it won't fit in the uint128 + // representation for printing. + const double small_exact_value = + 0.000000000000000000000000000000000000752316384526264005099991383822237233803945956334136013765601092018187046051025390625; // NOLINT + assert(small_exact_value == std::pow(2, -120)); + // Round up at a 5xx. + EXPECT_EQ(format("%.37f", small_exact_value), + "0.0000000000000000000000000000000000008"); + // Round down at a <5 + EXPECT_EQ(format("%.38f", small_exact_value), + "0.00000000000000000000000000000000000075"); + // Round up at a >5 + EXPECT_EQ(format("%.41f", small_exact_value), + "0.00000000000000000000000000000000000075232"); + // Nine handling + EXPECT_EQ(format("%.55f", small_exact_value), + "0.0000000000000000000000000000000000007523163845262640051"); + EXPECT_EQ(format("%.56f", small_exact_value), + "0.00000000000000000000000000000000000075231638452626400510"); + EXPECT_EQ(format("%.57f", small_exact_value), + "0.000000000000000000000000000000000000752316384526264005100"); + EXPECT_EQ(format("%.58f", small_exact_value), + "0.0000000000000000000000000000000000007523163845262640051000"); + // Round down the last nine + EXPECT_EQ(format("%.59f", small_exact_value), + "0.00000000000000000000000000000000000075231638452626400509999"); + // Round up the last nine + EXPECT_EQ(format("%.79f", small_exact_value), + "0.000000000000000000000000000000000000" + "7523163845262640050999913838222372338039460"); + + // Round to even (down) + EXPECT_EQ(format("%.119f", small_exact_value), + "0.000000000000000000000000000000000000" + "75231638452626400509999138382223723380" + "394595633413601376560109201818704605102539062"); + // Exact + EXPECT_EQ(format("%.120f", small_exact_value), + "0.000000000000000000000000000000000000" + "75231638452626400509999138382223723380" + "3945956334136013765601092018187046051025390625"); + // Round to even (up), let make the last digits 75 instead of 25 + EXPECT_EQ(format("%.119f", small_exact_value + std::pow(2, -119)), + "0.000000000000000000000000000000000002" + "25694915357879201529997415146671170141" + "183786900240804129680327605456113815307617188"); + // Exact, just to check. + EXPECT_EQ(format("%.120f", small_exact_value + std::pow(2, -119)), + "0.000000000000000000000000000000000002" + "25694915357879201529997415146671170141" + "1837869002408041296803276054561138153076171875"); +} + +TEST_F(FormatConvertTest, DoubleRoundA) { + const NativePrintfTraits &native_traits = VerifyNativeImplementation(); + std::string s; + const auto format = [&](const char *fmt, double d) -> std::string & { + s.clear(); + FormatArgImpl args[1] = {FormatArgImpl(d)}; + AppendPack(&s, UntypedFormatSpecImpl(fmt), absl::MakeSpan(args)); + if (native_traits.hex_float_has_glibc_rounding) { + EXPECT_EQ(StrPrint(fmt, d), s); + } + return s; + }; + + // 0x1.00018000p+100 + const double on_boundary_odd = 1267679614447900152596896153600.0; + EXPECT_EQ(format("%.0a", on_boundary_odd), "0x1p+100"); + EXPECT_EQ(format("%.1a", on_boundary_odd), "0x1.0p+100"); + EXPECT_EQ(format("%.2a", on_boundary_odd), "0x1.00p+100"); + EXPECT_EQ(format("%.3a", on_boundary_odd), "0x1.000p+100"); + EXPECT_EQ(format("%.4a", on_boundary_odd), "0x1.0002p+100"); // round + EXPECT_EQ(format("%.5a", on_boundary_odd), "0x1.00018p+100"); + EXPECT_EQ(format("%.6a", on_boundary_odd), "0x1.000180p+100"); + + // 0x1.00028000p-2 + const double on_boundary_even = 0.250009536743164062500; + EXPECT_EQ(format("%.0a", on_boundary_even), "0x1p-2"); + EXPECT_EQ(format("%.1a", on_boundary_even), "0x1.0p-2"); + EXPECT_EQ(format("%.2a", on_boundary_even), "0x1.00p-2"); + EXPECT_EQ(format("%.3a", on_boundary_even), "0x1.000p-2"); + EXPECT_EQ(format("%.4a", on_boundary_even), "0x1.0002p-2"); // no round + EXPECT_EQ(format("%.5a", on_boundary_even), "0x1.00028p-2"); + EXPECT_EQ(format("%.6a", on_boundary_even), "0x1.000280p-2"); + + // 0x1.00018001p+1 + const double slightly_over = 2.00004577683284878730773925781250; + EXPECT_EQ(format("%.0a", slightly_over), "0x1p+1"); + EXPECT_EQ(format("%.1a", slightly_over), "0x1.0p+1"); + EXPECT_EQ(format("%.2a", slightly_over), "0x1.00p+1"); + EXPECT_EQ(format("%.3a", slightly_over), "0x1.000p+1"); + EXPECT_EQ(format("%.4a", slightly_over), "0x1.0002p+1"); + EXPECT_EQ(format("%.5a", slightly_over), "0x1.00018p+1"); + EXPECT_EQ(format("%.6a", slightly_over), "0x1.000180p+1"); + + // 0x1.00017fffp+0 + const double slightly_under = 1.000022887950763106346130371093750; + EXPECT_EQ(format("%.0a", slightly_under), "0x1p+0"); + EXPECT_EQ(format("%.1a", slightly_under), "0x1.0p+0"); + EXPECT_EQ(format("%.2a", slightly_under), "0x1.00p+0"); + EXPECT_EQ(format("%.3a", slightly_under), "0x1.000p+0"); + EXPECT_EQ(format("%.4a", slightly_under), "0x1.0001p+0"); + EXPECT_EQ(format("%.5a", slightly_under), "0x1.00018p+0"); + EXPECT_EQ(format("%.6a", slightly_under), "0x1.000180p+0"); + EXPECT_EQ(format("%.7a", slightly_under), "0x1.0001800p+0"); + + // 0x1.1b3829ac28058p+3 + const double hex_value = 8.85060580848964661981881363317370414733886718750; + EXPECT_EQ(format("%.0a", hex_value), "0x1p+3"); + EXPECT_EQ(format("%.1a", hex_value), "0x1.2p+3"); + EXPECT_EQ(format("%.2a", hex_value), "0x1.1bp+3"); + EXPECT_EQ(format("%.3a", hex_value), "0x1.1b4p+3"); + EXPECT_EQ(format("%.4a", hex_value), "0x1.1b38p+3"); + EXPECT_EQ(format("%.5a", hex_value), "0x1.1b383p+3"); + EXPECT_EQ(format("%.6a", hex_value), "0x1.1b382ap+3"); + EXPECT_EQ(format("%.7a", hex_value), "0x1.1b3829bp+3"); + EXPECT_EQ(format("%.8a", hex_value), "0x1.1b3829acp+3"); + EXPECT_EQ(format("%.9a", hex_value), "0x1.1b3829ac3p+3"); + EXPECT_EQ(format("%.10a", hex_value), "0x1.1b3829ac28p+3"); + EXPECT_EQ(format("%.11a", hex_value), "0x1.1b3829ac280p+3"); + EXPECT_EQ(format("%.12a", hex_value), "0x1.1b3829ac2806p+3"); + EXPECT_EQ(format("%.13a", hex_value), "0x1.1b3829ac28058p+3"); + EXPECT_EQ(format("%.14a", hex_value), "0x1.1b3829ac280580p+3"); + EXPECT_EQ(format("%.15a", hex_value), "0x1.1b3829ac2805800p+3"); + EXPECT_EQ(format("%.16a", hex_value), "0x1.1b3829ac28058000p+3"); + EXPECT_EQ(format("%.17a", hex_value), "0x1.1b3829ac280580000p+3"); + EXPECT_EQ(format("%.18a", hex_value), "0x1.1b3829ac2805800000p+3"); + EXPECT_EQ(format("%.19a", hex_value), "0x1.1b3829ac28058000000p+3"); + EXPECT_EQ(format("%.20a", hex_value), "0x1.1b3829ac280580000000p+3"); + EXPECT_EQ(format("%.21a", hex_value), "0x1.1b3829ac2805800000000p+3"); + + // 0x1.0818283848586p+3 + const double hex_value2 = 8.2529488658208371987257123691961169242858886718750; + EXPECT_EQ(format("%.0a", hex_value2), "0x1p+3"); + EXPECT_EQ(format("%.1a", hex_value2), "0x1.1p+3"); + EXPECT_EQ(format("%.2a", hex_value2), "0x1.08p+3"); + EXPECT_EQ(format("%.3a", hex_value2), "0x1.082p+3"); + EXPECT_EQ(format("%.4a", hex_value2), "0x1.0818p+3"); + EXPECT_EQ(format("%.5a", hex_value2), "0x1.08183p+3"); + EXPECT_EQ(format("%.6a", hex_value2), "0x1.081828p+3"); + EXPECT_EQ(format("%.7a", hex_value2), "0x1.0818284p+3"); + EXPECT_EQ(format("%.8a", hex_value2), "0x1.08182838p+3"); + EXPECT_EQ(format("%.9a", hex_value2), "0x1.081828385p+3"); + EXPECT_EQ(format("%.10a", hex_value2), "0x1.0818283848p+3"); + EXPECT_EQ(format("%.11a", hex_value2), "0x1.08182838486p+3"); + EXPECT_EQ(format("%.12a", hex_value2), "0x1.081828384858p+3"); + EXPECT_EQ(format("%.13a", hex_value2), "0x1.0818283848586p+3"); + EXPECT_EQ(format("%.14a", hex_value2), "0x1.08182838485860p+3"); + EXPECT_EQ(format("%.15a", hex_value2), "0x1.081828384858600p+3"); + EXPECT_EQ(format("%.16a", hex_value2), "0x1.0818283848586000p+3"); + EXPECT_EQ(format("%.17a", hex_value2), "0x1.08182838485860000p+3"); + EXPECT_EQ(format("%.18a", hex_value2), "0x1.081828384858600000p+3"); + EXPECT_EQ(format("%.19a", hex_value2), "0x1.0818283848586000000p+3"); + EXPECT_EQ(format("%.20a", hex_value2), "0x1.08182838485860000000p+3"); + EXPECT_EQ(format("%.21a", hex_value2), "0x1.081828384858600000000p+3"); +} + +TEST_F(FormatConvertTest, LongDoubleRoundA) { + if (std::numeric_limits<long double>::digits % 4 != 0) { + // This test doesn't really make sense to run on platforms where a long + // double has a different mantissa size (mod 4) than Prod, since then the + // leading digit will be formatted differently. + return; + } + const NativePrintfTraits &native_traits = VerifyNativeImplementation(); + std::string s; + const auto format = [&](const char *fmt, long double d) -> std::string & { + s.clear(); + FormatArgImpl args[1] = {FormatArgImpl(d)}; + AppendPack(&s, UntypedFormatSpecImpl(fmt), absl::MakeSpan(args)); + if (native_traits.hex_float_has_glibc_rounding && + native_traits.hex_float_optimizes_leading_digit_bit_count) { + EXPECT_EQ(StrPrint(fmt, d), s); + } + return s; + }; + + // 0x8.8p+4 + const long double on_boundary_even = 136.0; + EXPECT_EQ(format("%.0La", on_boundary_even), "0x8p+4"); + EXPECT_EQ(format("%.1La", on_boundary_even), "0x8.8p+4"); + EXPECT_EQ(format("%.2La", on_boundary_even), "0x8.80p+4"); + EXPECT_EQ(format("%.3La", on_boundary_even), "0x8.800p+4"); + EXPECT_EQ(format("%.4La", on_boundary_even), "0x8.8000p+4"); + EXPECT_EQ(format("%.5La", on_boundary_even), "0x8.80000p+4"); + EXPECT_EQ(format("%.6La", on_boundary_even), "0x8.800000p+4"); + + // 0x9.8p+4 + const long double on_boundary_odd = 152.0; + EXPECT_EQ(format("%.0La", on_boundary_odd), "0xap+4"); + EXPECT_EQ(format("%.1La", on_boundary_odd), "0x9.8p+4"); + EXPECT_EQ(format("%.2La", on_boundary_odd), "0x9.80p+4"); + EXPECT_EQ(format("%.3La", on_boundary_odd), "0x9.800p+4"); + EXPECT_EQ(format("%.4La", on_boundary_odd), "0x9.8000p+4"); + EXPECT_EQ(format("%.5La", on_boundary_odd), "0x9.80000p+4"); + EXPECT_EQ(format("%.6La", on_boundary_odd), "0x9.800000p+4"); + + // 0x8.80001p+24 + const long double slightly_over = 142606352.0; + EXPECT_EQ(format("%.0La", slightly_over), "0x9p+24"); + EXPECT_EQ(format("%.1La", slightly_over), "0x8.8p+24"); + EXPECT_EQ(format("%.2La", slightly_over), "0x8.80p+24"); + EXPECT_EQ(format("%.3La", slightly_over), "0x8.800p+24"); + EXPECT_EQ(format("%.4La", slightly_over), "0x8.8000p+24"); + EXPECT_EQ(format("%.5La", slightly_over), "0x8.80001p+24"); + EXPECT_EQ(format("%.6La", slightly_over), "0x8.800010p+24"); + + // 0x8.7ffffp+24 + const long double slightly_under = 142606320.0; + EXPECT_EQ(format("%.0La", slightly_under), "0x8p+24"); + EXPECT_EQ(format("%.1La", slightly_under), "0x8.8p+24"); + EXPECT_EQ(format("%.2La", slightly_under), "0x8.80p+24"); + EXPECT_EQ(format("%.3La", slightly_under), "0x8.800p+24"); + EXPECT_EQ(format("%.4La", slightly_under), "0x8.8000p+24"); + EXPECT_EQ(format("%.5La", slightly_under), "0x8.7ffffp+24"); + EXPECT_EQ(format("%.6La", slightly_under), "0x8.7ffff0p+24"); + EXPECT_EQ(format("%.7La", slightly_under), "0x8.7ffff00p+24"); + + // 0xc.0828384858688000p+128 + const long double eights = 4094231060438608800781871108094404067328.0; + EXPECT_EQ(format("%.0La", eights), "0xcp+128"); + EXPECT_EQ(format("%.1La", eights), "0xc.1p+128"); + EXPECT_EQ(format("%.2La", eights), "0xc.08p+128"); + EXPECT_EQ(format("%.3La", eights), "0xc.083p+128"); + EXPECT_EQ(format("%.4La", eights), "0xc.0828p+128"); + EXPECT_EQ(format("%.5La", eights), "0xc.08284p+128"); + EXPECT_EQ(format("%.6La", eights), "0xc.082838p+128"); + EXPECT_EQ(format("%.7La", eights), "0xc.0828385p+128"); + EXPECT_EQ(format("%.8La", eights), "0xc.08283848p+128"); + EXPECT_EQ(format("%.9La", eights), "0xc.082838486p+128"); + EXPECT_EQ(format("%.10La", eights), "0xc.0828384858p+128"); + EXPECT_EQ(format("%.11La", eights), "0xc.08283848587p+128"); + EXPECT_EQ(format("%.12La", eights), "0xc.082838485868p+128"); + EXPECT_EQ(format("%.13La", eights), "0xc.0828384858688p+128"); + EXPECT_EQ(format("%.14La", eights), "0xc.08283848586880p+128"); + EXPECT_EQ(format("%.15La", eights), "0xc.082838485868800p+128"); + EXPECT_EQ(format("%.16La", eights), "0xc.0828384858688000p+128"); +} + +// We don't actually store the results. This is just to exercise the rest of the +// machinery. +struct NullSink { + friend void AbslFormatFlush(NullSink *sink, string_view str) {} +}; + +template <typename... T> +bool FormatWithNullSink(absl::string_view fmt, const T &... a) { + NullSink sink; + FormatArgImpl args[] = {FormatArgImpl(a)...}; + return FormatUntyped(&sink, UntypedFormatSpecImpl(fmt), absl::MakeSpan(args)); +} + +TEST_F(FormatConvertTest, ExtremeWidthPrecision) { + for (const char *fmt : {"f"}) { + for (double d : {1e-100, 1.0, 1e100}) { + constexpr int max = std::numeric_limits<int>::max(); + EXPECT_TRUE(FormatWithNullSink(std::string("%.*") + fmt, max, d)); + EXPECT_TRUE(FormatWithNullSink(std::string("%1.*") + fmt, max, d)); + EXPECT_TRUE(FormatWithNullSink(std::string("%*") + fmt, max, d)); + EXPECT_TRUE(FormatWithNullSink(std::string("%*.*") + fmt, max, max, d)); + } + } } TEST_F(FormatConvertTest, LongDouble) { - const char *const kFormats[] = {"%", "%.3", "%8.5", "%9", +#ifdef _MSC_VER + // MSVC has a different rounding policy than us so we can't test our + // implementation against the native one there. + return; +#endif // _MSC_VER + const NativePrintfTraits &native_traits = VerifyNativeImplementation(); + const char *const kFormats[] = {"%", "%.3", "%8.5", "%9", "%.5000", "%.60", "%+", "% ", "%-10"}; - // This value is not representable in double, but it is in long double that - // uses the extended format. - // This is to verify that we are not truncating the value mistakenly through a - // double. - long double very_precise = 10000000000000000.25L; - std::vector<long double> doubles = { 0.0, -0.0, - very_precise, - 1 / very_precise, std::numeric_limits<long double>::max(), -std::numeric_limits<long double>::max(), std::numeric_limits<long double>::min(), @@ -556,28 +1074,65 @@ TEST_F(FormatConvertTest, LongDouble) { std::numeric_limits<long double>::infinity(), -std::numeric_limits<long double>::infinity()}; + for (long double base : {1.L, 12.L, 123.L, 1234.L, 12345.L, 123456.L, + 1234567.L, 12345678.L, 123456789.L, 1234567890.L, + 12345678901.L, 123456789012.L, 1234567890123.L, + // This value is not representable in double, but it + // is in long double that uses the extended format. + // This is to verify that we are not truncating the + // value mistakenly through a double. + 10000000000000000.25L}) { + for (int exp : {-1000, -500, 0, 500, 1000}) { + for (int sign : {1, -1}) { + doubles.push_back(sign * std::ldexp(base, exp)); + doubles.push_back(sign / std::ldexp(base, exp)); + } + } + } + + // Regression tests + // + // Using a string literal because not all platforms support hex literals or it + // might be out of range. + doubles.push_back(std::strtold("-0xf.ffffffb5feafffbp-16324L", nullptr)); + for (const char *fmt : kFormats) { for (char f : {'f', 'F', // 'g', 'G', // 'a', 'A', // 'e', 'E'}) { std::string fmt_str = std::string(fmt) + 'L' + f; + + if (fmt == absl::string_view("%.5000") && f != 'f' && f != 'F' && + f != 'a' && f != 'A') { + // This particular test takes way too long with snprintf. + // Disable for the case we are not implementing natively. + continue; + } + + if (f == 'a' || f == 'A') { + if (!native_traits.hex_float_has_glibc_rounding || + !native_traits.hex_float_optimizes_leading_digit_bit_count) { + continue; + } + } + for (auto d : doubles) { FormatArgImpl arg(d); UntypedFormatSpecImpl format(fmt_str); // We use ASSERT_EQ here because failures are usually correlated and a // bug would print way too many failed expectations causing the test to // time out. - ASSERT_EQ(StrPrint(fmt_str.c_str(), d), - FormatPack(format, {&arg, 1})) + ASSERT_EQ(StrPrint(fmt_str.c_str(), d), FormatPack(format, {&arg, 1})) << fmt_str << " " << StrPrint("%.18Lg", d) << " " - << StrPrint("%.999Lf", d); + << StrPrint("%La", d) << " " << StrPrint("%.1080Lf", d); } } } } -TEST_F(FormatConvertTest, IntAsFloat) { +TEST_F(FormatConvertTest, IntAsDouble) { + const NativePrintfTraits &native_traits = VerifyNativeImplementation(); const int kMin = std::numeric_limits<int>::min(); const int kMax = std::numeric_limits<int>::max(); const int ia[] = { @@ -593,14 +1148,17 @@ TEST_F(FormatConvertTest, IntAsFloat) { const char *fmt; }; const double dx = static_cast<double>(fx); - const Expectation kExpect[] = { - { __LINE__, StrPrint("%f", dx), "%f" }, - { __LINE__, StrPrint("%12f", dx), "%12f" }, - { __LINE__, StrPrint("%.12f", dx), "%.12f" }, - { __LINE__, StrPrint("%12a", dx), "%12a" }, - { __LINE__, StrPrint("%.12a", dx), "%.12a" }, + std::vector<Expectation> expect = { + {__LINE__, StrPrint("%f", dx), "%f"}, + {__LINE__, StrPrint("%12f", dx), "%12f"}, + {__LINE__, StrPrint("%.12f", dx), "%.12f"}, + {__LINE__, StrPrint("%.12a", dx), "%.12a"}, }; - for (const Expectation &e : kExpect) { + if (native_traits.hex_float_uses_minimal_precision_when_not_specified) { + Expectation ex = {__LINE__, StrPrint("%12a", dx), "%12a"}; + expect.push_back(ex); + } + for (const Expectation &e : expect) { SCOPED_TRACE(e.line); SCOPED_TRACE(e.fmt); UntypedFormatSpecImpl format(e.fmt); @@ -645,6 +1203,25 @@ TEST_F(FormatConvertTest, ExpectedFailures) { EXPECT_TRUE(FormatFails("%*d", "")); } +// Sanity check to make sure that we are testing what we think we're testing on +// e.g. the x86_64+glibc platform. +TEST_F(FormatConvertTest, GlibcHasCorrectTraits) { +#if !defined(__GLIBC__) || !defined(__x86_64__) + return; +#endif + const NativePrintfTraits &native_traits = VerifyNativeImplementation(); + // If one of the following tests break then it is either because the above PP + // macro guards failed to exclude a new platform (likely) or because something + // has changed in the implemention of glibc sprintf float formatting behavior. + // If the latter, then the code that computes these flags needs to be + // revisited and/or possibly the StrFormat implementation. + EXPECT_TRUE(native_traits.hex_float_has_glibc_rounding); + EXPECT_TRUE(native_traits.hex_float_prefers_denormal_repr); + EXPECT_TRUE( + native_traits.hex_float_uses_minimal_precision_when_not_specified); + EXPECT_TRUE(native_traits.hex_float_optimizes_leading_digit_bit_count); +} + } // namespace } // namespace str_format_internal ABSL_NAMESPACE_END diff --git a/absl/strings/internal/str_format/extension.cc b/absl/strings/internal/str_format/extension.cc index 2e5bc2ce..bb0d96cf 100644 --- a/absl/strings/internal/str_format/extension.cc +++ b/absl/strings/internal/str_format/extension.cc @@ -33,16 +33,40 @@ std::string Flags::ToString() const { return s; } -bool FormatSinkImpl::PutPaddedString(string_view v, int w, int p, bool l) { +#define ABSL_INTERNAL_X_VAL(id) \ + constexpr absl::FormatConversionChar FormatConversionCharInternal::id; +ABSL_INTERNAL_CONVERSION_CHARS_EXPAND_(ABSL_INTERNAL_X_VAL, ) +#undef ABSL_INTERNAL_X_VAL +// NOLINTNEXTLINE(readability-redundant-declaration) +constexpr absl::FormatConversionChar FormatConversionCharInternal::kNone; + +#define ABSL_INTERNAL_CHAR_SET_CASE(c) \ + constexpr FormatConversionCharSet FormatConversionCharSetInternal::c; +ABSL_INTERNAL_CONVERSION_CHARS_EXPAND_(ABSL_INTERNAL_CHAR_SET_CASE, ) +#undef ABSL_INTERNAL_CHAR_SET_CASE + +// NOLINTNEXTLINE(readability-redundant-declaration) +constexpr FormatConversionCharSet FormatConversionCharSetInternal::kStar; +// NOLINTNEXTLINE(readability-redundant-declaration) +constexpr FormatConversionCharSet FormatConversionCharSetInternal::kIntegral; +// NOLINTNEXTLINE(readability-redundant-declaration) +constexpr FormatConversionCharSet FormatConversionCharSetInternal::kFloating; +// NOLINTNEXTLINE(readability-redundant-declaration) +constexpr FormatConversionCharSet FormatConversionCharSetInternal::kNumeric; +// NOLINTNEXTLINE(readability-redundant-declaration) +constexpr FormatConversionCharSet FormatConversionCharSetInternal::kPointer; + +bool FormatSinkImpl::PutPaddedString(string_view value, int width, + int precision, bool left) { size_t space_remaining = 0; - if (w >= 0) space_remaining = w; - size_t n = v.size(); - if (p >= 0) n = std::min(n, static_cast<size_t>(p)); - string_view shown(v.data(), n); + if (width >= 0) space_remaining = width; + size_t n = value.size(); + if (precision >= 0) n = std::min(n, static_cast<size_t>(precision)); + string_view shown(value.data(), n); space_remaining = Excess(shown.size(), space_remaining); - if (!l) Append(space_remaining, ' '); + if (!left) Append(space_remaining, ' '); Append(shown); - if (l) Append(space_remaining, ' '); + if (left) Append(space_remaining, ' '); return true; } diff --git a/absl/strings/internal/str_format/extension.h b/absl/strings/internal/str_format/extension.h index d1665753..a9b9e137 100644 --- a/absl/strings/internal/str_format/extension.h +++ b/absl/strings/internal/str_format/extension.h @@ -24,11 +24,16 @@ #include "absl/base/config.h" #include "absl/base/port.h" +#include "absl/meta/type_traits.h" #include "absl/strings/internal/str_format/output.h" #include "absl/strings/string_view.h" namespace absl { ABSL_NAMESPACE_BEGIN + +enum class FormatConversionChar : uint8_t; +enum class FormatConversionCharSet : uint64_t; + namespace str_format_internal { class FormatRawSinkImpl { @@ -102,7 +107,7 @@ class FormatSinkImpl { size_t size() const { return size_; } // Put 'v' to 'sink' with specified width, precision, and left flag. - bool PutPaddedString(string_view v, int w, int p, bool l); + bool PutPaddedString(string_view v, int width, int precision, bool left); template <typename T> T Wrap() { @@ -139,7 +144,7 @@ struct Flags { // clang-format off #define ABSL_INTERNAL_CONVERSION_CHARS_EXPAND_(X_VAL, X_SEP) \ /* text */ \ - X_VAL(c) X_SEP X_VAL(C) X_SEP X_VAL(s) X_SEP X_VAL(S) X_SEP \ + X_VAL(c) X_SEP X_VAL(s) X_SEP \ /* ints */ \ X_VAL(d) X_SEP X_VAL(i) X_SEP X_VAL(o) X_SEP \ X_VAL(u) X_SEP X_VAL(x) X_SEP X_VAL(X) X_SEP \ @@ -148,14 +153,39 @@ struct Flags { X_VAL(g) X_SEP X_VAL(G) X_SEP X_VAL(a) X_SEP X_VAL(A) X_SEP \ /* misc */ \ X_VAL(n) X_SEP X_VAL(p) +// clang-format on + +// This type should not be referenced, it exists only to provide labels +// internally that match the values declared in FormatConversionChar in +// str_format.h. This is meant to allow internal libraries to use the same +// declared interface type as the public interface +// (absl::StrFormatConversionChar) while keeping the definition in a public +// header. +// Internal libraries should use the form +// `FormatConversionCharInternal::c`, `FormatConversionCharInternal::kNone` for +// comparisons. Use in switch statements is not recommended due to a bug in how +// gcc 4.9 -Wswitch handles declared but undefined enums. +struct FormatConversionCharInternal { + FormatConversionCharInternal() = delete; -enum class FormatConversionChar : uint8_t { - c, C, s, S, // text + private: + // clang-format off + enum class Enum : uint8_t { + c, s, // text d, i, o, u, x, X, // int f, F, e, E, g, G, a, A, // float n, p, // misc - kNone, - none = kNone + kNone + }; + // clang-format on + public: +#define ABSL_INTERNAL_X_VAL(id) \ + static constexpr FormatConversionChar id = \ + static_cast<FormatConversionChar>(Enum::id); + ABSL_INTERNAL_CONVERSION_CHARS_EXPAND_(ABSL_INTERNAL_X_VAL, ) +#undef ABSL_INTERNAL_X_VAL + static constexpr FormatConversionChar kNone = + static_cast<FormatConversionChar>(Enum::kNone); }; // clang-format on @@ -163,95 +193,56 @@ inline FormatConversionChar FormatConversionCharFromChar(char c) { switch (c) { #define ABSL_INTERNAL_X_VAL(id) \ case #id[0]: \ - return FormatConversionChar::id; + return FormatConversionCharInternal::id; ABSL_INTERNAL_CONVERSION_CHARS_EXPAND_(ABSL_INTERNAL_X_VAL, ) #undef ABSL_INTERNAL_X_VAL } - return FormatConversionChar::kNone; -} - -inline int FormatConversionCharRadix(FormatConversionChar c) { - switch (c) { - case FormatConversionChar::x: - case FormatConversionChar::X: - case FormatConversionChar::a: - case FormatConversionChar::A: - case FormatConversionChar::p: - return 16; - case FormatConversionChar::o: - return 8; - default: - return 10; - } + return FormatConversionCharInternal::kNone; } inline bool FormatConversionCharIsUpper(FormatConversionChar c) { - switch (c) { - case FormatConversionChar::X: - case FormatConversionChar::F: - case FormatConversionChar::E: - case FormatConversionChar::G: - case FormatConversionChar::A: - return true; - default: - return false; - } -} - -inline bool FormatConversionCharIsSigned(FormatConversionChar c) { - switch (c) { - case FormatConversionChar::d: - case FormatConversionChar::i: - return true; - default: - return false; - } -} - -inline bool FormatConversionCharIsIntegral(FormatConversionChar c) { - switch (c) { - case FormatConversionChar::d: - case FormatConversionChar::i: - case FormatConversionChar::u: - case FormatConversionChar::o: - case FormatConversionChar::x: - case FormatConversionChar::X: - return true; - default: - return false; + if (c == FormatConversionCharInternal::X || + c == FormatConversionCharInternal::F || + c == FormatConversionCharInternal::E || + c == FormatConversionCharInternal::G || + c == FormatConversionCharInternal::A) { + return true; + } else { + return false; } } inline bool FormatConversionCharIsFloat(FormatConversionChar c) { - switch (c) { - case FormatConversionChar::a: - case FormatConversionChar::e: - case FormatConversionChar::f: - case FormatConversionChar::g: - case FormatConversionChar::A: - case FormatConversionChar::E: - case FormatConversionChar::F: - case FormatConversionChar::G: - return true; - default: - return false; + if (c == FormatConversionCharInternal::a || + c == FormatConversionCharInternal::e || + c == FormatConversionCharInternal::f || + c == FormatConversionCharInternal::g || + c == FormatConversionCharInternal::A || + c == FormatConversionCharInternal::E || + c == FormatConversionCharInternal::F || + c == FormatConversionCharInternal::G) { + return true; + } else { + return false; } } inline char FormatConversionCharToChar(FormatConversionChar c) { - switch (c) { -#define ABSL_INTERNAL_X_VAL(e) \ - case FormatConversionChar::e: \ + if (c == FormatConversionCharInternal::kNone) { + return '\0'; + +#define ABSL_INTERNAL_X_VAL(e) \ + } else if (c == FormatConversionCharInternal::e) { \ return #e[0]; #define ABSL_INTERNAL_X_SEP - ABSL_INTERNAL_CONVERSION_CHARS_EXPAND_(ABSL_INTERNAL_X_VAL, - ABSL_INTERNAL_X_SEP) - case FormatConversionChar::kNone: - return '\0'; + ABSL_INTERNAL_CONVERSION_CHARS_EXPAND_(ABSL_INTERNAL_X_VAL, + ABSL_INTERNAL_X_SEP) + } else { + return '\0'; + } + #undef ABSL_INTERNAL_X_VAL #undef ABSL_INTERNAL_X_SEP - } - return '\0'; } // The associated char. @@ -263,7 +254,7 @@ inline std::ostream& operator<<(std::ostream& os, FormatConversionChar v) { struct FormatConversionSpecImplFriend; -class FormatConversionSpec { +class FormatConversionSpecImpl { public: // Width and precison are not specified, no flags are set. bool is_basic() const { return flags_.basic; } @@ -276,7 +267,7 @@ class FormatConversionSpec { FormatConversionChar conversion_char() const { // Keep this field first in the struct . It generates better code when // accessing it when ConversionSpec is passed by value in registers. - static_assert(offsetof(FormatConversionSpec, conv_) == 0, ""); + static_assert(offsetof(FormatConversionSpecImpl, conv_) == 0, ""); return conv_; } @@ -287,41 +278,65 @@ class FormatConversionSpec { // negative value. int precision() const { return precision_; } - // Deprecated (use has_x_flag() instead). - Flags flags() const { return flags_; } - // Deprecated - FormatConversionChar conv() const { return conversion_char(); } + template <typename T> + T Wrap() { + return T(*this); + } private: friend struct str_format_internal::FormatConversionSpecImplFriend; - FormatConversionChar conv_ = FormatConversionChar::kNone; + FormatConversionChar conv_ = FormatConversionCharInternal::kNone; Flags flags_; int width_; int precision_; }; struct FormatConversionSpecImplFriend final { - static void SetFlags(Flags f, FormatConversionSpec* conv) { + static void SetFlags(Flags f, FormatConversionSpecImpl* conv) { conv->flags_ = f; } static void SetConversionChar(FormatConversionChar c, - FormatConversionSpec* conv) { + FormatConversionSpecImpl* conv) { conv->conv_ = c; } - static void SetWidth(int w, FormatConversionSpec* conv) { conv->width_ = w; } - static void SetPrecision(int p, FormatConversionSpec* conv) { + static void SetWidth(int w, FormatConversionSpecImpl* conv) { + conv->width_ = w; + } + static void SetPrecision(int p, FormatConversionSpecImpl* conv) { conv->precision_ = p; } - static std::string FlagsToString(const FormatConversionSpec& spec) { + static std::string FlagsToString(const FormatConversionSpecImpl& spec) { return spec.flags_.ToString(); } }; -constexpr uint64_t FormatConversionCharToConvValue(char conv) { +// Type safe OR operator. +// We need this for two reasons: +// 1. operator| on enums makes them decay to integers and the result is an +// integer. We need the result to stay as an enum. +// 2. We use "enum class" which would not work even if we accepted the decay. +constexpr FormatConversionCharSet FormatConversionCharSetUnion( + FormatConversionCharSet a) { + return a; +} + +template <typename... CharSet> +constexpr FormatConversionCharSet FormatConversionCharSetUnion( + FormatConversionCharSet a, CharSet... rest) { + return static_cast<FormatConversionCharSet>( + static_cast<uint64_t>(a) | + static_cast<uint64_t>(FormatConversionCharSetUnion(rest...))); +} + +constexpr uint64_t FormatConversionCharToConvInt(FormatConversionChar c) { + return uint64_t{1} << (1 + static_cast<uint8_t>(c)); +} + +constexpr uint64_t FormatConversionCharToConvInt(char conv) { return -#define ABSL_INTERNAL_CHAR_SET_CASE(c) \ - conv == #c[0] \ - ? (uint64_t{1} << (1 + static_cast<uint8_t>(FormatConversionChar::c))) \ +#define ABSL_INTERNAL_CHAR_SET_CASE(c) \ + conv == #c[0] \ + ? FormatConversionCharToConvInt(FormatConversionCharInternal::c) \ : ABSL_INTERNAL_CONVERSION_CHARS_EXPAND_(ABSL_INTERNAL_CHAR_SET_CASE, ) #undef ABSL_INTERNAL_CHAR_SET_CASE @@ -330,28 +345,29 @@ constexpr uint64_t FormatConversionCharToConvValue(char conv) { : 0; } -enum class FormatConversionCharSet : uint64_t { -#define ABSL_INTERNAL_CHAR_SET_CASE(c) \ - c = FormatConversionCharToConvValue(#c[0]), +constexpr FormatConversionCharSet FormatConversionCharToConvValue(char conv) { + return static_cast<FormatConversionCharSet>( + FormatConversionCharToConvInt(conv)); +} + +struct FormatConversionCharSetInternal { +#define ABSL_INTERNAL_CHAR_SET_CASE(c) \ + static constexpr FormatConversionCharSet c = \ + FormatConversionCharToConvValue(#c[0]); ABSL_INTERNAL_CONVERSION_CHARS_EXPAND_(ABSL_INTERNAL_CHAR_SET_CASE, ) #undef ABSL_INTERNAL_CHAR_SET_CASE // Used for width/precision '*' specification. - kStar = FormatConversionCharToConvValue('*'), - // Some predefined values: - kIntegral = d | i | u | o | x | X, - kFloating = a | e | f | g | A | E | F | G, - kNumeric = kIntegral | kFloating, - kString = s, - kPointer = p, - - // The following are deprecated - star = kStar, - integral = kIntegral, - floating = kFloating, - numeric = kNumeric, - string = kString, - pointer = kPointer + static constexpr FormatConversionCharSet kStar = + FormatConversionCharToConvValue('*'); + + static constexpr FormatConversionCharSet kIntegral = + FormatConversionCharSetUnion(d, i, u, o, x, X); + static constexpr FormatConversionCharSet kFloating = + FormatConversionCharSetUnion(a, e, f, g, A, E, F, G); + static constexpr FormatConversionCharSet kNumeric = + FormatConversionCharSetUnion(kIntegral, kFloating); + static constexpr FormatConversionCharSet kPointer = p; }; // Type safe OR operator. @@ -361,18 +377,29 @@ enum class FormatConversionCharSet : uint64_t { // 2. We use "enum class" which would not work even if we accepted the decay. constexpr FormatConversionCharSet operator|(FormatConversionCharSet a, FormatConversionCharSet b) { - return FormatConversionCharSet(static_cast<uint64_t>(a) | - static_cast<uint64_t>(b)); + return FormatConversionCharSetUnion(a, b); +} + +// Overloaded conversion functions to support absl::ParsedFormat. +// Get a conversion with a single character in it. +constexpr FormatConversionCharSet ToFormatConversionCharSet(char c) { + return static_cast<FormatConversionCharSet>( + FormatConversionCharToConvValue(c)); } // Get a conversion with a single character in it. -constexpr FormatConversionCharSet ConversionCharToConv(char c) { - return FormatConversionCharSet(FormatConversionCharToConvValue(c)); +constexpr FormatConversionCharSet ToFormatConversionCharSet( + FormatConversionCharSet c) { + return c; } +template <typename T> +void ToFormatConversionCharSet(T) = delete; + // Checks whether `c` exists in `set`. constexpr bool Contains(FormatConversionCharSet set, char c) { - return (static_cast<uint64_t>(set) & FormatConversionCharToConvValue(c)) != 0; + return (static_cast<uint64_t>(set) & + static_cast<uint64_t>(FormatConversionCharToConvValue(c))) != 0; } // Checks whether all the characters in `c` are contained in `set` @@ -382,31 +409,16 @@ constexpr bool Contains(FormatConversionCharSet set, static_cast<uint64_t>(c); } -// Return type of the AbslFormatConvert() functions. -// The FormatConversionCharSet template parameter is used to inform the -// framework of what conversion characters are supported by that -// AbslFormatConvert routine. -template <FormatConversionCharSet C> -struct FormatConvertResult { - static constexpr FormatConversionCharSet kConv = C; - bool value; -}; - -template <FormatConversionCharSet C> -constexpr FormatConversionCharSet FormatConvertResult<C>::kConv; +// Checks whether all the characters in `c` are contained in `set` +constexpr bool Contains(FormatConversionCharSet set, FormatConversionChar c) { + return (static_cast<uint64_t>(set) & FormatConversionCharToConvInt(c)) != 0; +} // Return capacity - used, clipped to a minimum of 0. inline size_t Excess(size_t used, size_t capacity) { return used < capacity ? capacity - used : 0; } -// Type alias for use during migration. -using ConversionChar = FormatConversionChar; -using ConversionSpec = FormatConversionSpec; -using Conv = FormatConversionCharSet; -template <FormatConversionCharSet C> -using ConvertResult = FormatConvertResult<C>; - } // namespace str_format_internal ABSL_NAMESPACE_END diff --git a/absl/strings/internal/str_format/extension_test.cc b/absl/strings/internal/str_format/extension_test.cc index 4e23fefb..1c93fdb1 100644 --- a/absl/strings/internal/str_format/extension_test.cc +++ b/absl/strings/internal/str_format/extension_test.cc @@ -19,9 +19,26 @@ #include <random> #include <string> +#include "gtest/gtest.h" #include "absl/strings/str_format.h" +#include "absl/strings/string_view.h" -#include "gtest/gtest.h" +namespace my_namespace { +class UserDefinedType { + public: + UserDefinedType() = default; + + void Append(absl::string_view str) { value_.append(str.data(), str.size()); } + const std::string& Value() const { return value_; } + + friend void AbslFormatFlush(UserDefinedType* x, absl::string_view str) { + x->Append(str); + } + + private: + std::string value_; +}; +} // namespace my_namespace namespace { @@ -63,4 +80,19 @@ TEST(FormatExtensionTest, SinkAppendChars) { EXPECT_EQ(actual, expected); } } + +TEST(FormatExtensionTest, VerifyEnumEquality) { +#define X_VAL(id) \ + EXPECT_EQ(absl::FormatConversionChar::id, \ + absl::str_format_internal::FormatConversionCharInternal::id); + ABSL_INTERNAL_CONVERSION_CHARS_EXPAND_(X_VAL, ); +#undef X_VAL + +#define X_VAL(id) \ + EXPECT_EQ(absl::FormatConversionCharSet::id, \ + absl::str_format_internal::FormatConversionCharSetInternal::id); + ABSL_INTERNAL_CONVERSION_CHARS_EXPAND_(X_VAL, ); +#undef X_VAL +} + } // namespace diff --git a/absl/strings/internal/str_format/float_conversion.cc b/absl/strings/internal/str_format/float_conversion.cc index d4c647c3..20aeada5 100644 --- a/absl/strings/internal/str_format/float_conversion.cc +++ b/absl/strings/internal/str_format/float_conversion.cc @@ -1,12 +1,23 @@ #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 @@ -14,13 +25,901 @@ namespace str_format_internal { namespace { -char *CopyStringTo(string_view v, char *out) { +// 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( + 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 ConversionSpec &conv, +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; @@ -33,17 +932,17 @@ bool FallbackToSnprintf(const Float v, const ConversionSpec &conv, if (std::is_same<long double, Float>()) { *fp++ = 'L'; } - *fp++ = FormatConversionCharToChar(conv.conv()); + *fp++ = FormatConversionCharToChar(conv.conversion_char()); *fp = 0; assert(fp < fmt + sizeof(fmt)); } std::string space(512, '\0'); - string_view result; + 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 = string_view(space.data(), n); + result = absl::string_view(space.data(), n); break; } space.resize(n + 1); @@ -96,21 +995,24 @@ enum class FormatStyle { Fixed, Precision }; // Otherwise, return false. template <typename Float> bool ConvertNonNumericFloats(char sign_char, Float v, - const ConversionSpec &conv, FormatSinkImpl *sink) { + const FormatConversionSpecImpl &conv, + FormatSinkImpl *sink) { char text[4], *ptr = text; - if (sign_char) *ptr++ = sign_char; + if (sign_char != '\0') *ptr++ = sign_char; if (std::isnan(v)) { - ptr = std::copy_n(FormatConversionCharIsUpper(conv.conv()) ? "NAN" : "nan", - 3, ptr); + ptr = std::copy_n( + FormatConversionCharIsUpper(conv.conversion_char()) ? "NAN" : "nan", 3, + ptr); } else if (std::isinf(v)) { - ptr = std::copy_n(FormatConversionCharIsUpper(conv.conv()) ? "INF" : "inf", - 3, ptr); + 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.flags().left); + conv.has_left_flag()); } // Round up the last digit of the value. @@ -170,7 +1072,12 @@ constexpr bool CanFitMantissa() { template <typename Float> struct Decomposed { - Float mantissa; + 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; }; @@ -181,7 +1088,8 @@ Decomposed<Float> Decompose(Float v) { Float m = std::frexp(v, &exp); m = std::ldexp(m, std::numeric_limits<Float>::digits); exp -= std::numeric_limits<Float>::digits; - return {m, exp}; + + return {static_cast<typename Decomposed<Float>::MantissaType>(m), exp}; } // Print 'digits' as decimal. @@ -350,31 +1258,32 @@ bool FloatToBuffer(Decomposed<Float> decomposed, int precision, Buffer *out, return false; } -void WriteBufferToSink(char sign_char, string_view str, - const ConversionSpec &conv, FormatSinkImpl *sink) { +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.flags().left) { + if (conv.has_left_flag()) { right_spaces = missing_chars; - } else if (conv.flags().zero) { + } else if (conv.has_zero_flag()) { zeros = missing_chars; } else { left_spaces = missing_chars; } sink->Append(left_spaces, ' '); - if (sign_char) sink->Append(1, sign_char); + 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 ConversionSpec &conv, +bool FloatToSink(const Float v, const FormatConversionSpecImpl &conv, FormatSinkImpl *sink) { // Print the sign or the sign column. Float abs_v = v; @@ -382,9 +1291,9 @@ bool FloatToSink(const Float v, const ConversionSpec &conv, if (std::signbit(abs_v)) { sign_char = '-'; abs_v = -abs_v; - } else if (conv.flags().show_pos) { + } else if (conv.has_show_pos_flag()) { sign_char = '+'; - } else if (conv.flags().sign_col) { + } else if (conv.has_sign_col_flag()) { sign_char = ' '; } @@ -401,89 +1310,92 @@ bool FloatToSink(const Float v, const ConversionSpec &conv, Buffer buffer; - switch (conv.conv()) { - case ConversionChar::f: - case ConversionChar::F: - if (!FloatToBuffer<FormatStyle::Fixed>(decomposed, precision, &buffer, - nullptr)) { - return FallbackToSnprintf(v, conv, sink); - } - if (!conv.flags().alt && buffer.back() == '.') buffer.pop_back(); - break; - - case ConversionChar::e: - case ConversionChar::E: - if (!FloatToBuffer<FormatStyle::Precision>(decomposed, precision, &buffer, - &exp)) { - return FallbackToSnprintf(v, conv, sink); - } - if (!conv.flags().alt && buffer.back() == '.') buffer.pop_back(); - PrintExponent(exp, FormatConversionCharIsUpper(conv.conv()) ? 'E' : 'e', - &buffer); - break; - - case ConversionChar::g: - case ConversionChar::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.flags().alt) { - while (buffer.back() == '0') buffer.pop_back(); - if (buffer.back() == '.') buffer.pop_back(); - } - if (exp) { - PrintExponent(exp, FormatConversionCharIsUpper(conv.conv()) ? 'E' : 'e', - &buffer); - } - break; + FormatConversionChar c = conv.conversion_char(); - case ConversionChar::a: - case ConversionChar::A: + 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); - - default: - return false; + } + 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, - string_view(buffer.begin, buffer.end - buffer.begin), conv, - sink); + absl::string_view(buffer.begin, buffer.end - buffer.begin), + conv, sink); return true; } } // namespace -bool ConvertFloatImpl(long double v, const ConversionSpec &conv, +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 ConversionSpec &conv, +bool ConvertFloatImpl(float v, const FormatConversionSpecImpl &conv, FormatSinkImpl *sink) { - return FloatToSink(v, conv, sink); + return FloatToSink(static_cast<double>(v), conv, sink); } -bool ConvertFloatImpl(double v, const ConversionSpec &conv, +bool ConvertFloatImpl(double v, const FormatConversionSpecImpl &conv, FormatSinkImpl *sink) { return FloatToSink(v, conv, sink); } diff --git a/absl/strings/internal/str_format/float_conversion.h b/absl/strings/internal/str_format/float_conversion.h index 49a6a636..e78bc191 100644 --- a/absl/strings/internal/str_format/float_conversion.h +++ b/absl/strings/internal/str_format/float_conversion.h @@ -7,13 +7,13 @@ namespace absl { ABSL_NAMESPACE_BEGIN namespace str_format_internal { -bool ConvertFloatImpl(float v, const ConversionSpec &conv, +bool ConvertFloatImpl(float v, const FormatConversionSpecImpl &conv, FormatSinkImpl *sink); -bool ConvertFloatImpl(double v, const ConversionSpec &conv, +bool ConvertFloatImpl(double v, const FormatConversionSpecImpl &conv, FormatSinkImpl *sink); -bool ConvertFloatImpl(long double v, const ConversionSpec &conv, +bool ConvertFloatImpl(long double v, const FormatConversionSpecImpl &conv, FormatSinkImpl *sink); } // namespace str_format_internal diff --git a/absl/strings/internal/str_format/output.h b/absl/strings/internal/str_format/output.h index 28b288b7..8030dae0 100644 --- a/absl/strings/internal/str_format/output.h +++ b/absl/strings/internal/str_format/output.h @@ -30,9 +30,6 @@ namespace absl { ABSL_NAMESPACE_BEGIN - -class Cord; - namespace str_format_internal { // RawSink implementation that writes into a char* buffer. @@ -77,12 +74,6 @@ inline void AbslFormatFlush(std::ostream* out, string_view s) { out->write(s.data(), s.size()); } -template <class AbslCord, typename = typename std::enable_if< - std::is_same<AbslCord, absl::Cord>::value>::type> -inline void AbslFormatFlush(AbslCord* out, string_view s) { - out->Append(s); -} - inline void AbslFormatFlush(FILERawSink* sink, string_view v) { sink->Write(v); } @@ -91,10 +82,11 @@ inline void AbslFormatFlush(BufferRawSink* sink, string_view v) { sink->Write(v); } +// This is a SFINAE to get a better compiler error message when the type +// is not supported. template <typename T> -auto InvokeFlush(T* out, string_view s) - -> decltype(str_format_internal::AbslFormatFlush(out, s)) { - str_format_internal::AbslFormatFlush(out, s); +auto InvokeFlush(T* out, string_view s) -> decltype(AbslFormatFlush(out, s)) { + AbslFormatFlush(out, s); } } // namespace str_format_internal diff --git a/absl/strings/internal/str_format/output_test.cc b/absl/strings/internal/str_format/output_test.cc index e54e6f70..ce2e91a0 100644 --- a/absl/strings/internal/str_format/output_test.cc +++ b/absl/strings/internal/str_format/output_test.cc @@ -19,6 +19,7 @@ #include "gmock/gmock.h" #include "gtest/gtest.h" +#include "absl/strings/cord.h" namespace absl { ABSL_NAMESPACE_BEGIN @@ -37,6 +38,12 @@ TEST(InvokeFlush, Stream) { EXPECT_EQ(str.str(), "ABCDEF"); } +TEST(InvokeFlush, Cord) { + absl::Cord str("ABC"); + str_format_internal::InvokeFlush(&str, "DEF"); + EXPECT_EQ(str, "ABCDEF"); +} + TEST(BufferRawSink, Limits) { char buf[16]; { @@ -70,4 +77,3 @@ TEST(BufferRawSink, Limits) { } // namespace ABSL_NAMESPACE_END } // namespace absl - diff --git a/absl/strings/internal/str_format/parser.cc b/absl/strings/internal/str_format/parser.cc index aab68db9..cc55dfa9 100644 --- a/absl/strings/internal/str_format/parser.cc +++ b/absl/strings/internal/str_format/parser.cc @@ -17,7 +17,7 @@ namespace absl { ABSL_NAMESPACE_BEGIN namespace str_format_internal { -using CC = ConversionChar; +using CC = FormatConversionCharInternal; using LM = LengthMod; ABSL_CONST_INIT const ConvTag kTags[256] = { @@ -29,9 +29,9 @@ ABSL_CONST_INIT const ConvTag kTags[256] = { {}, {}, {}, {}, {}, {}, {}, {}, // 28-2f {}, {}, {}, {}, {}, {}, {}, {}, // 30-37 {}, {}, {}, {}, {}, {}, {}, {}, // 38-3f - {}, CC::A, {}, CC::C, {}, CC::E, CC::F, CC::G, // @ABCDEFG + {}, CC::A, {}, {}, {}, CC::E, CC::F, CC::G, // @ABCDEFG {}, {}, {}, {}, LM::L, {}, {}, {}, // HIJKLMNO - {}, {}, {}, CC::S, {}, {}, {}, {}, // PQRSTUVW + {}, {}, {}, {}, {}, {}, {}, {}, // PQRSTUVW CC::X, {}, {}, {}, {}, {}, {}, {}, // XYZ[\]^_ {}, CC::a, {}, CC::c, CC::d, CC::e, CC::f, CC::g, // `abcdefg LM::h, CC::i, LM::j, {}, LM::l, {}, CC::n, CC::o, // hijklmno @@ -296,15 +296,17 @@ struct ParsedFormatBase::ParsedFormatConsumer { char* data_pos; }; -ParsedFormatBase::ParsedFormatBase(string_view format, bool allow_ignored, - std::initializer_list<Conv> convs) +ParsedFormatBase::ParsedFormatBase( + string_view format, bool allow_ignored, + std::initializer_list<FormatConversionCharSet> convs) : data_(format.empty() ? nullptr : new char[format.size()]) { has_error_ = !ParseFormatString(format, ParsedFormatConsumer(this)) || !MatchesConversions(allow_ignored, convs); } bool ParsedFormatBase::MatchesConversions( - bool allow_ignored, std::initializer_list<Conv> convs) const { + bool allow_ignored, + std::initializer_list<FormatConversionCharSet> convs) const { std::unordered_set<int> used; auto add_if_valid_conv = [&](int pos, char c) { if (static_cast<size_t>(pos) > convs.size() || diff --git a/absl/strings/internal/str_format/parser.h b/absl/strings/internal/str_format/parser.h index 45c90d1d..fffed04f 100644 --- a/absl/strings/internal/str_format/parser.h +++ b/absl/strings/internal/str_format/parser.h @@ -67,7 +67,7 @@ struct UnboundConversion { Flags flags; LengthMod length_mod = LengthMod::none; - ConversionChar conv = FormatConversionChar::kNone; + FormatConversionChar conv = FormatConversionCharInternal::kNone; }; // Consume conversion spec prefix (not including '%') of [p, end) if valid. @@ -83,7 +83,7 @@ const char* ConsumeUnboundConversion(const char* p, const char* end, // conversions. class ConvTag { public: - constexpr ConvTag(ConversionChar conversion_char) // NOLINT + constexpr ConvTag(FormatConversionChar conversion_char) // NOLINT : tag_(static_cast<int8_t>(conversion_char)) {} // We invert the length modifiers to make them negative so that we can easily // test for them. @@ -94,9 +94,9 @@ class ConvTag { bool is_conv() const { return tag_ >= 0; } bool is_length() const { return tag_ < 0 && tag_ != -128; } - ConversionChar as_conv() const { + FormatConversionChar as_conv() const { assert(is_conv()); - return static_cast<ConversionChar>(tag_); + return static_cast<FormatConversionChar>(tag_); } LengthMod as_length() const { assert(is_length()); @@ -143,7 +143,7 @@ bool ParseFormatString(string_view src, Consumer consumer) { auto tag = GetTagForChar(percent[1]); if (tag.is_conv()) { if (ABSL_PREDICT_FALSE(next_arg < 0)) { - // This indicates an error in the format std::string. + // This indicates an error in the format string. // The only way to get `next_arg < 0` here is to have a positional // argument first which sets next_arg to -1 and then a non-positional // argument. @@ -186,8 +186,9 @@ constexpr bool EnsureConstexpr(string_view s) { class ParsedFormatBase { public: - explicit ParsedFormatBase(string_view format, bool allow_ignored, - std::initializer_list<Conv> convs); + explicit ParsedFormatBase( + string_view format, bool allow_ignored, + std::initializer_list<FormatConversionCharSet> convs); ParsedFormatBase(const ParsedFormatBase& other) { *this = other; } @@ -234,8 +235,9 @@ class ParsedFormatBase { private: // Returns whether the conversions match and if !allow_ignored it verifies // that all conversions are used by the format. - bool MatchesConversions(bool allow_ignored, - std::initializer_list<Conv> convs) const; + bool MatchesConversions( + bool allow_ignored, + std::initializer_list<FormatConversionCharSet> convs) const; struct ParsedFormatConsumer; @@ -280,14 +282,14 @@ class ParsedFormatBase { // This is the only API that allows the user to pass a runtime specified format // string. These factory functions will return NULL if the format does not match // the conversions requested by the user. -template <str_format_internal::Conv... C> +template <FormatConversionCharSet... C> class ExtendedParsedFormat : public str_format_internal::ParsedFormatBase { public: explicit ExtendedParsedFormat(string_view format) #ifdef ABSL_INTERNAL_ENABLE_FORMAT_CHECKER __attribute__(( enable_if(str_format_internal::EnsureConstexpr(format), - "Format std::string is not constexpr."), + "Format string is not constexpr."), enable_if(str_format_internal::ValidFormatImpl<C...>(format), "Format specified does not match the template arguments."))) #endif // ABSL_INTERNAL_ENABLE_FORMAT_CHECKER diff --git a/absl/strings/internal/str_format/parser_test.cc b/absl/strings/internal/str_format/parser_test.cc index 1b1ee030..5aced987 100644 --- a/absl/strings/internal/str_format/parser_test.cc +++ b/absl/strings/internal/str_format/parser_test.cc @@ -41,23 +41,23 @@ TEST(LengthModTest, Names) { TEST(ConversionCharTest, Names) { struct Expectation { - ConversionChar id; + FormatConversionChar id; char name; }; // clang-format off const Expectation kExpect[] = { -#define X(c) {ConversionChar::c, #c[0]} - X(c), X(C), X(s), X(S), // text +#define X(c) {FormatConversionCharInternal::c, #c[0]} + X(c), X(s), // text X(d), X(i), X(o), X(u), X(x), X(X), // int X(f), X(F), X(e), X(E), X(g), X(G), X(a), X(A), // float X(n), X(p), // misc #undef X - {ConversionChar::none, '\0'}, + {FormatConversionCharInternal::kNone, '\0'}, }; // clang-format on for (auto e : kExpect) { SCOPED_TRACE(e.name); - ConversionChar v = e.id; + FormatConversionChar v = e.id; EXPECT_EQ(e.name, FormatConversionCharToChar(v)); } } @@ -349,7 +349,8 @@ TEST_F(ParsedFormatTest, ValueSemantics) { ParsedFormatBase p2 = p1; // copy construct (empty) EXPECT_EQ(SummarizeParsedFormat(p1), SummarizeParsedFormat(p2)); - p1 = ParsedFormatBase("hello%s", true, {Conv::s}); // move assign + p1 = ParsedFormatBase("hello%s", true, + {FormatConversionCharSetInternal::s}); // move assign EXPECT_EQ("[hello]{s:1$s}", SummarizeParsedFormat(p1)); ParsedFormatBase p3 = p1; // copy construct (nonempty) @@ -367,7 +368,7 @@ TEST_F(ParsedFormatTest, ValueSemantics) { struct ExpectParse { const char* in; - std::initializer_list<Conv> conv_set; + std::initializer_list<FormatConversionCharSet> conv_set; const char* out; }; @@ -377,9 +378,9 @@ TEST_F(ParsedFormatTest, Parsing) { const ExpectParse kExpect[] = { {"", {}, ""}, {"ab", {}, "[ab]"}, - {"a%d", {Conv::d}, "[a]{d:1$d}"}, - {"a%+d", {Conv::d}, "[a]{+d:1$d}"}, - {"a% d", {Conv::d}, "[a]{ d:1$d}"}, + {"a%d", {FormatConversionCharSetInternal::d}, "[a]{d:1$d}"}, + {"a%+d", {FormatConversionCharSetInternal::d}, "[a]{+d:1$d}"}, + {"a% d", {FormatConversionCharSetInternal::d}, "[a]{ d:1$d}"}, {"a%b %d", {}, "[a]!"}, // stop after error }; for (const auto& e : kExpect) { @@ -391,13 +392,13 @@ TEST_F(ParsedFormatTest, Parsing) { TEST_F(ParsedFormatTest, ParsingFlagOrder) { const ExpectParse kExpect[] = { - {"a%+ 0d", {Conv::d}, "[a]{+ 0d:1$d}"}, - {"a%+0 d", {Conv::d}, "[a]{+0 d:1$d}"}, - {"a%0+ d", {Conv::d}, "[a]{0+ d:1$d}"}, - {"a% +0d", {Conv::d}, "[a]{ +0d:1$d}"}, - {"a%0 +d", {Conv::d}, "[a]{0 +d:1$d}"}, - {"a% 0+d", {Conv::d}, "[a]{ 0+d:1$d}"}, - {"a%+ 0+d", {Conv::d}, "[a]{+ 0+d:1$d}"}, + {"a%+ 0d", {FormatConversionCharSetInternal::d}, "[a]{+ 0d:1$d}"}, + {"a%+0 d", {FormatConversionCharSetInternal::d}, "[a]{+0 d:1$d}"}, + {"a%0+ d", {FormatConversionCharSetInternal::d}, "[a]{0+ d:1$d}"}, + {"a% +0d", {FormatConversionCharSetInternal::d}, "[a]{ +0d:1$d}"}, + {"a%0 +d", {FormatConversionCharSetInternal::d}, "[a]{0 +d:1$d}"}, + {"a% 0+d", {FormatConversionCharSetInternal::d}, "[a]{ 0+d:1$d}"}, + {"a%+ 0+d", {FormatConversionCharSetInternal::d}, "[a]{+ 0+d:1$d}"}, }; for (const auto& e : kExpect) { SCOPED_TRACE(e.in); |