diff options
Diffstat (limited to 'absl/strings/internal')
| -rw-r--r-- | absl/strings/internal/str_format/arg.cc | 112 | ||||
| -rw-r--r-- | absl/strings/internal/str_format/arg.h | 78 | ||||
| -rw-r--r-- | absl/strings/internal/str_format/arg_test.cc | 5 | ||||
| -rw-r--r-- | absl/strings/internal/str_format/bind.h | 7 | ||||
| -rw-r--r-- | absl/strings/internal/str_format/checker_test.cc | 2 | ||||
| -rw-r--r-- | absl/strings/internal/str_format/convert_test.cc | 245 | ||||
| -rw-r--r-- | absl/strings/internal/str_format/extension.h | 5 | ||||
| -rw-r--r-- | absl/strings/internal/str_format/float_conversion.cc | 694 | ||||
| -rw-r--r-- | absl/strings/internal/str_format/float_conversion.h | 6 | ||||
| -rw-r--r-- | absl/strings/internal/str_format/parser.h | 8 | ||||
| -rw-r--r-- | absl/strings/internal/str_format/parser_test.cc | 6 |
11 files changed, 1014 insertions, 154 deletions
diff --git a/absl/strings/internal/str_format/arg.cc b/absl/strings/internal/str_format/arg.cc index a112071c43a7..964f25f7da64 100644 --- a/absl/strings/internal/str_format/arg.cc +++ b/absl/strings/internal/str_format/arg.cc @@ -167,24 +167,26 @@ class IntDigits { // 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 IntDigits &as_digits, - const ConversionSpec conv) { + const FormatConversionSpecImpl conv) { // always show 0x for %p. - bool alt = conv.has_alt_flag() || conv.conversion_char() == ConversionChar::p; - bool hex = (conv.conversion_char() == FormatConversionChar::x || - conv.conversion_char() == FormatConversionChar::X || - conv.conversion_char() == FormatConversionChar::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 && hex && !as_digits.without_neg_or_zero().empty()) { - return conv.conversion_char() == FormatConversionChar::X ? "0X" : "0x"; + return conv.conversion_char() == FormatConversionCharInternal::X ? "0X" + : "0x"; } return {}; } -string_view SignColumn(bool neg, const ConversionSpec conv) { - if (conv.conversion_char() == FormatConversionChar::d || - conv.conversion_char() == FormatConversionChar::i) { +string_view SignColumn(bool neg, const FormatConversionSpecImpl conv) { + if (conv.conversion_char() == FormatConversionCharInternal::d || + conv.conversion_char() == FormatConversionCharInternal::i) { if (neg) return "-"; if (conv.has_show_pos_flag()) return "+"; if (conv.has_sign_col_flag()) return " "; @@ -192,7 +194,7 @@ string_view SignColumn(bool neg, const ConversionSpec conv) { 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(); @@ -204,7 +206,8 @@ bool ConvertCharImpl(unsigned char v, const ConversionSpec conv, } bool ConvertIntImplInnerSlow(const IntDigits &as_digits, - const ConversionSpec conv, FormatSinkImpl *sink) { + const FormatConversionSpecImpl conv, + FormatSinkImpl *sink) { // Print as a sequence of Substrings: // [left_spaces][sign][base_indicator][zeroes][formatted][right_spaces] size_t fill = 0; @@ -224,7 +227,8 @@ bool ConvertIntImplInnerSlow(const IntDigits &as_digits, if (!precision_specified) precision = 1; - if (conv.has_alt_flag() && conv.conversion_char() == 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." @@ -258,42 +262,43 @@ bool ConvertIntImplInnerSlow(const IntDigits &as_digits, } template <typename T> -bool ConvertIntArg(T v, const ConversionSpec conv, FormatSinkImpl *sink) { +bool ConvertIntArg(T v, const FormatConversionSpecImpl conv, + FormatSinkImpl *sink) { using U = typename MakeUnsigned<T>::type; IntDigits as_digits; switch (conv.conversion_char()) { - case FormatConversionChar::c: + case FormatConversionCharInternal::c: return ConvertCharImpl(static_cast<unsigned char>(v), conv, sink); - case FormatConversionChar::o: + case FormatConversionCharInternal::o: as_digits.PrintAsOct(static_cast<U>(v)); break; - case FormatConversionChar::x: + case FormatConversionCharInternal::x: as_digits.PrintAsHexLower(static_cast<U>(v)); break; - case FormatConversionChar::X: + case FormatConversionCharInternal::X: as_digits.PrintAsHexUpper(static_cast<U>(v)); break; - case FormatConversionChar::u: + case FormatConversionCharInternal::u: as_digits.PrintAsDec(static_cast<U>(v)); break; - case FormatConversionChar::d: - case FormatConversionChar::i: + case FormatConversionCharInternal::d: + case FormatConversionCharInternal::i: as_digits.PrintAsDec(v); break; - case FormatConversionChar::a: - case FormatConversionChar::e: - case FormatConversionChar::f: - case FormatConversionChar::g: - case FormatConversionChar::A: - case FormatConversionChar::E: - case FormatConversionChar::F: - case FormatConversionChar::G: + case FormatConversionCharInternal::a: + case FormatConversionCharInternal::e: + case FormatConversionCharInternal::f: + case FormatConversionCharInternal::g: + case FormatConversionCharInternal::A: + case FormatConversionCharInternal::E: + case FormatConversionCharInternal::F: + case FormatConversionCharInternal::G: return ConvertFloatImpl(static_cast<double>(v), conv, sink); default: @@ -308,12 +313,13 @@ bool ConvertIntArg(T v, const ConversionSpec conv, FormatSinkImpl *sink) { } template <typename T> -bool ConvertFloatArg(T v, const ConversionSpec conv, FormatSinkImpl *sink) { +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.conversion_char() != FormatConversionCharInternal::s) return false; if (conv.is_basic()) { @@ -328,19 +334,20 @@ inline bool ConvertStringArg(string_view v, const ConversionSpec conv, // ==================== Strings ==================== StringConvertResult FormatConvertImpl(const std::string &v, - const ConversionSpec conv, + const FormatConversionSpecImpl conv, FormatSinkImpl *sink) { return {ConvertStringArg(v, conv, sink)}; } -StringConvertResult FormatConvertImpl(string_view v, const ConversionSpec conv, +StringConvertResult FormatConvertImpl(string_view v, + const FormatConversionSpecImpl conv, FormatSinkImpl *sink) { return {ConvertStringArg(v, conv, sink)}; } ArgConvertResult<FormatConversionCharSetUnion( FormatConversionCharSetInternal::s, FormatConversionCharSetInternal::p)> -FormatConvertImpl(const char *v, const ConversionSpec conv, +FormatConvertImpl(const char *v, const FormatConversionSpecImpl conv, FormatSinkImpl *sink) { if (conv.conversion_char() == FormatConversionCharInternal::p) return {FormatConvertImpl(VoidPtr(v), conv, sink).value}; @@ -358,7 +365,7 @@ FormatConvertImpl(const char *v, const ConversionSpec conv, // ==================== Raw pointers ==================== ArgConvertResult<FormatConversionCharSetInternal::p> FormatConvertImpl( - VoidPtr v, const ConversionSpec conv, FormatSinkImpl *sink) { + VoidPtr v, const FormatConversionSpecImpl conv, FormatSinkImpl *sink) { if (conv.conversion_char() != FormatConversionCharInternal::p) return {false}; if (!v.value) { sink->Append("(nil)"); @@ -370,82 +377,87 @@ ArgConvertResult<FormatConversionCharSetInternal::p> FormatConvertImpl( } // ==================== 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 f4ac940ab43f..9a1e5ef276e8 100644 --- a/absl/strings/internal/str_format/arg.h +++ b/absl/strings/internal/str_format/arg.h @@ -67,20 +67,24 @@ constexpr FormatConversionCharSet ExtractCharSet(ArgConvertResult<C>) { using StringConvertResult = ArgConvertResult<FormatConversionCharSetInternal::s>; ArgConvertResult<FormatConversionCharSetInternal::p> FormatConvertImpl( - VoidPtr v, ConversionSpec conv, FormatSinkImpl* sink); + VoidPtr v, FormatConversionSpecImpl conv, FormatSinkImpl* sink); // Strings. -StringConvertResult FormatConvertImpl(const std::string& v, ConversionSpec conv, +StringConvertResult FormatConvertImpl(const std::string& v, + FormatConversionSpecImpl conv, FormatSinkImpl* sink); -StringConvertResult FormatConvertImpl(string_view v, ConversionSpec conv, +StringConvertResult FormatConvertImpl(string_view v, + FormatConversionSpecImpl conv, FormatSinkImpl* sink); ArgConvertResult<FormatConversionCharSetUnion( FormatConversionCharSetInternal::s, FormatConversionCharSetInternal::p)> -FormatConvertImpl(const char* v, ConversionSpec conv, FormatSinkImpl* sink); +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, - ConversionSpec conv, + FormatConversionSpecImpl conv, FormatSinkImpl* sink) { if (conv.conversion_char() != FormatConversionCharInternal::s) { return {false}; @@ -127,50 +131,55 @@ 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); } @@ -181,11 +190,11 @@ 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> StringConvertResult FormatConvertImpl(const StreamedWrapper<T>& v, - ConversionSpec conv, + FormatConversionSpecImpl conv, FormatSinkImpl* out) { std::ostringstream oss; oss << v.v_; @@ -198,7 +207,8 @@ StringConvertResult FormatConvertImpl(const StreamedWrapper<T>& v, struct FormatCountCaptureHelper { template <class T = int> static ArgConvertResult<FormatConversionCharSetInternal::n> ConvertHelper( - const FormatCountCapture& v, ConversionSpec conv, FormatSinkImpl* sink) { + const FormatCountCapture& v, FormatConversionSpecImpl conv, + FormatSinkImpl* sink) { const absl::enable_if_t<sizeof(T) != 0, FormatCountCapture>& v2 = v; if (conv.conversion_char() != @@ -212,7 +222,8 @@ struct FormatCountCaptureHelper { template <class T = int> ArgConvertResult<FormatConversionCharSetInternal::n> FormatConvertImpl( - const FormatCountCapture& v, ConversionSpec conv, FormatSinkImpl* sink) { + const FormatCountCapture& v, FormatConversionSpecImpl conv, + FormatSinkImpl* sink) { return FormatCountCaptureHelper::ConvertHelper(v, conv, sink); } @@ -221,13 +232,13 @@ ArgConvertResult<FormatConversionCharSetInternal::n> FormatConvertImpl( 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); } @@ -251,7 +262,7 @@ 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 @@ -393,7 +404,7 @@ 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.conversion_char() == FormatConversionCharInternal::kNone)) { @@ -410,8 +421,9 @@ 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 8d30d8b8ce34..37e5b7545f84 100644 --- a/absl/strings/internal/str_format/arg_test.cc +++ b/absl/strings/internal/str_format/arg_test.cc @@ -95,8 +95,9 @@ 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(FormatConversionChar::s, + &conv); FormatConversionSpecImplFriend::SetFlags(Flags(), &conv); FormatConversionSpecImplFriend::SetWidth(-1, &conv); FormatConversionSpecImplFriend::SetPrecision(-1, &conv); diff --git a/absl/strings/internal/str_format/bind.h b/absl/strings/internal/str_format/bind.h index 05105d8d5c9c..585246e77e56 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; } @@ -119,7 +119,7 @@ class FormatSpecTemplate #endif // ABSL_INTERNAL_ENABLE_FORMAT_CHECKER - template <Conv... C, + template <FormatConversionCharSet... C, typename = typename std::enable_if< AllOf(sizeof...(C) == sizeof...(Args), Contains(Args, C)...)>::type> @@ -190,7 +190,8 @@ class StreamedWrapper { private: template <typename S> friend ArgConvertResult<FormatConversionCharSetInternal::s> FormatConvertImpl( - const StreamedWrapper<S>& v, ConversionSpec conv, FormatSinkImpl* out); + const StreamedWrapper<S>& v, FormatConversionSpecImpl conv, + FormatSinkImpl* out); const T& v_; }; diff --git a/absl/strings/internal/str_format/checker_test.cc b/absl/strings/internal/str_format/checker_test.cc index 49a24b4026c3..233481747bfd 100644 --- a/absl/strings/internal/str_format/checker_test.cc +++ b/absl/strings/internal/str_format/checker_test.cc @@ -24,7 +24,7 @@ std::string ConvToString(FormatConversionCharSet conv) { } 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 cbcd7caf4665..dd167f76e68f 100644 --- a/absl/strings/internal/str_format/convert_test.cc +++ b/absl/strings/internal/str_format/convert_test.cc @@ -1,14 +1,18 @@ #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/types/optional.h" namespace absl { ABSL_NAMESPACE_BEGIN @@ -57,7 +61,7 @@ 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]; @@ -98,11 +102,18 @@ 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; } @@ -471,8 +482,8 @@ TEST_F(FormatConvertTest, Float) { #endif // _MSC_VER const char *const kFormats[] = { - "%", "%.3", "%8.5", "%9", "%.60", "%.30", "%03", "%+", - "% ", "%-10", "%#15.3", "%#.0", "%.0", "%1$*2$", "%1$.*2$"}; + "%", "%.3", "%8.5", "%500", "%.5000", "%.60", "%.30", "%03", + "%+", "% ", "%-10", "%#15.3", "%#.0", "%.0", "%1$*2$", "%1$.*2$"}; std::vector<double> doubles = {0.0, -0.0, @@ -489,11 +500,6 @@ 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 +518,204 @@ TEST_F(FormatConvertTest, Float) { } } + // 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 + + // 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); + 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') { + // This particular test takes way too long with snprintf. + // Disable for the case we are not implementing natively. + continue; + } + 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); + + 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("%.1080f", d); + } } } } } +TEST_F(FormatConvertTest, FloatRound) { + 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, 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 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(), @@ -556,22 +723,44 @@ 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)); + } + } + } + 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') { + // This particular test takes way too long with snprintf. + // Disable for the case we are not implementing natively. + 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); } } } diff --git a/absl/strings/internal/str_format/extension.h b/absl/strings/internal/str_format/extension.h index f0cffe1e6d6c..33903df0be39 100644 --- a/absl/strings/internal/str_format/extension.h +++ b/absl/strings/internal/str_format/extension.h @@ -411,11 +411,6 @@ 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 = FormatConversionSpecImpl; -using Conv = FormatConversionCharSet; - class FormatConversionSpec { public: // Width and precison are not specified, no flags are set. diff --git a/absl/strings/internal/str_format/float_conversion.cc b/absl/strings/internal/str_format/float_conversion.cc index d6858cfffd95..cdccc86f551c 100644 --- a/absl/strings/internal/str_format/float_conversion.cc +++ b/absl/strings/internal/str_format/float_conversion.cc @@ -1,12 +1,22 @@ #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/types/optional.h" +#include "absl/types/span.h" namespace absl { ABSL_NAMESPACE_BEGIN @@ -14,13 +24,640 @@ 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 size_t 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) { + assert(-exp < 0); + assert(-exp >= std::numeric_limits<long double>::min_exponent - 128); + static_assert( + StackArray::kMaxCapacity >= + (128 - std::numeric_limits<long double>::min_exponent + 31) / 32, + ""); + StackArray::RunWithCapacity((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(int total_size, const FormatState &state) { + int missing_chars = std::max(state.conv.width() - total_size, 0); + 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(absl::string_view data, int trailing_zeros, + const FormatState &state) { + 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'); + return; + } + + auto padding = + ExtraWidthToPadding((state.sign_char != '\0' ? 1 : 0) + + static_cast<int>(data.size()) + trailing_zeros, + 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'); + state.sink->Append(data); + state.sink->Append(trailing_zeros, '0'); + 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(absl::string_view(integral_digits_start, size), + static_cast<int>(state.precision - (fractional_digits_end - + fractional_digits_start)), + state); +} + +// 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 int total_digits = + btd.TotalDigits() + (state.ShouldPrintDot() ? 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 int total_digits = + /* 0 */ 1 + (state.ShouldPrintDot() ? 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); +} + +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; @@ -38,12 +675,12 @@ bool FallbackToSnprintf(const Float v, const ConversionSpec &conv, 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); @@ -96,9 +733,10 @@ 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.conversion_char()) ? "NAN" : "nan", 3, @@ -172,7 +810,12 @@ 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; }; @@ -183,7 +826,8 @@ 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. @@ -352,8 +996,9 @@ 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()) - @@ -369,14 +1014,14 @@ void WriteBufferToSink(char sign_char, string_view str, } 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; @@ -407,11 +1052,9 @@ bool FloatToSink(const Float v, const ConversionSpec &conv, if (c == FormatConversionCharInternal::f || c == FormatConversionCharInternal::F) { - if (!FloatToBuffer<FormatStyle::Fixed>(decomposed, precision, &buffer, - nullptr)) { - return FallbackToSnprintf(v, conv, sink); - } - if (!conv.has_alt_flag() && buffer.back() == '.') buffer.pop_back(); + 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, @@ -462,25 +1105,32 @@ bool FloatToSink(const Float v, const ConversionSpec &conv, } 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); } -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 49a6a63630ee..e78bc19106ff 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/parser.h b/absl/strings/internal/str_format/parser.h index fd2dc9704507..fffed04fa072 100644 --- a/absl/strings/internal/str_format/parser.h +++ b/absl/strings/internal/str_format/parser.h @@ -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()); @@ -282,7 +282,7 @@ 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) diff --git a/absl/strings/internal/str_format/parser_test.cc b/absl/strings/internal/str_format/parser_test.cc index 26f5bec6bcff..dae2d20f593e 100644 --- a/absl/strings/internal/str_format/parser_test.cc +++ b/absl/strings/internal/str_format/parser_test.cc @@ -41,7 +41,7 @@ TEST(LengthModTest, Names) { TEST(ConversionCharTest, Names) { struct Expectation { - ConversionChar id; + FormatConversionChar id; char name; }; // clang-format off @@ -57,7 +57,7 @@ TEST(ConversionCharTest, Names) { // 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)); } } @@ -368,7 +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; }; |