From fe447768b741e6f4a1cb529f93174e6dc6f86bec Mon Sep 17 00:00:00 2001 From: Abseil Team Date: Tue, 26 Mar 2024 11:41:28 -0700 Subject: [PATCH] Revert integer-to-string conversion optimizations pending more thorough analysis PiperOrigin-RevId: 619261152 Change-Id: Id3409b326c52ace0fda42537e0b91dbb2d6a2287 --- absl/base/macros.h | 48 ++++ absl/strings/numbers.cc | 440 +++++++++-------------------------- absl/strings/numbers.h | 163 +------------ absl/strings/numbers_test.cc | 7 - absl/strings/str_cat.cc | 146 +----------- absl/strings/str_cat.h | 150 ++++++------ absl/strings/str_cat_test.cc | 46 ---- 7 files changed, 245 insertions(+), 755 deletions(-) diff --git a/absl/base/macros.h b/absl/base/macros.h index f33cd192..b318f116 100644 --- a/absl/base/macros.h +++ b/absl/base/macros.h @@ -138,4 +138,52 @@ ABSL_NAMESPACE_END #define ABSL_INTERNAL_RETHROW do {} while (false) #endif // ABSL_HAVE_EXCEPTIONS +// ABSL_DEPRECATE_AND_INLINE() +// +// Marks a function or type alias as deprecated and tags it to be picked up for +// automated refactoring by go/cpp-inliner. It can added to inline function +// definitions or type aliases. It should only be used within a header file. It +// differs from `ABSL_DEPRECATED` in the following ways: +// +// 1. New uses of the function or type will be discouraged via Tricorder +// warnings. +// 2. If enabled via `METADATA`, automated changes will be sent out inlining the +// functions's body or replacing the type where it is used. +// +// For example: +// +// ABSL_DEPRECATE_AND_INLINE() inline int OldFunc(int x) { +// return NewFunc(x, 0); +// } +// +// will mark `OldFunc` as deprecated, and the go/cpp-inliner service will +// replace calls to `OldFunc(x)` with calls to `NewFunc(x, 0)`. Once all calls +// to `OldFunc` have been replaced, `OldFunc` can be deleted. +// +// See go/cpp-inliner for more information. +// +// Note: go/cpp-inliner is Google-internal service for automated refactoring. +// While open-source users do not have access to this service, the macro is +// provided for compatibility, and so that users receive deprecation warnings. +#if ABSL_HAVE_CPP_ATTRIBUTE(deprecated) && \ + ABSL_HAVE_CPP_ATTRIBUTE(clang::annotate) +#define ABSL_DEPRECATE_AND_INLINE() [[deprecated, clang::annotate("inline-me")]] +#elif ABSL_HAVE_CPP_ATTRIBUTE(deprecated) +#define ABSL_DEPRECATE_AND_INLINE() [[deprecated]] +#else +#define ABSL_DEPRECATE_AND_INLINE() +#endif + +// Requires the compiler to prove that the size of the given object is at least +// the expected amount. +#if ABSL_HAVE_ATTRIBUTE(diagnose_if) && ABSL_HAVE_BUILTIN(__builtin_object_size) +#define ABSL_INTERNAL_NEED_MIN_SIZE(Obj, N) \ + __attribute__((diagnose_if(__builtin_object_size(Obj, 0) < N, \ + "object size provably too small " \ + "(this would corrupt memory)", \ + "error"))) +#else +#define ABSL_INTERNAL_NEED_MIN_SIZE(Obj, N) +#endif + #endif // ABSL_BASE_MACROS_H_ diff --git a/absl/strings/numbers.cc b/absl/strings/numbers.cc index 882c3a8b..b57d9e82 100644 --- a/absl/strings/numbers.cc +++ b/absl/strings/numbers.cc @@ -20,9 +20,7 @@ #include #include #include // for DBL_DIG and FLT_DIG -#include #include // for HUGE_VAL -#include #include #include #include @@ -30,7 +28,6 @@ #include #include #include // NOLINT(build/c++11) -#include #include #include "absl/base/attributes.h" @@ -159,71 +156,28 @@ constexpr uint32_t kTwoZeroBytes = 0x0101 * '0'; constexpr uint64_t kFourZeroBytes = 0x01010101 * '0'; constexpr uint64_t kEightZeroBytes = 0x0101010101010101ull * '0'; -template -constexpr T Pow(T base, uint32_t n) { - // Exponentiation by squaring - return static_cast((n > 1 ? Pow(base * base, n >> 1) : static_cast(1)) * - ((n & 1) ? base : static_cast(1))); -} - -// Given n, calculates C where the following holds for all 0 <= x < Pow(100, n): -// x / Pow(10, n) == x * C / Pow(2, n * 10) -// In other words, it allows us to divide by a power of 10 via a single -// multiplication and bit shifts, assuming the input will be smaller than the -// square of that power of 10. -template -constexpr T ComputePowerOf100DivisionCoefficient(uint32_t n) { - if (n > 4) { - // This doesn't work for large powers of 100, due to overflow - abort(); - } - T denom = 16 - 1; - T num = (denom + 1) - 10; - T gcd = 3; // Greatest common divisor of numerator and denominator - denom = Pow(denom / gcd, n); - num = Pow(num / gcd, 9 * n); - T quotient = num / denom; - if (num % denom >= denom / 2) { - // Round up, since the remainder is more than half the denominator - ++quotient; - } - return quotient; -} - -// * kDivisionBy10Mul / kDivisionBy10Div is a division by 10 for values from 0 -// to 99. It's also a division of a structure [k takes 2 bytes][m takes 2 -// bytes], then * kDivisionBy10Mul / kDivisionBy10Div will be [k / 10][m / 10]. -// It allows parallel division. -constexpr uint64_t kDivisionBy10Mul = - ComputePowerOf100DivisionCoefficient(1); -static_assert(kDivisionBy10Mul == 103, - "division coefficient for 10 is incorrect"); +// * 103 / 1024 is a division by 10 for values from 0 to 99. It's also a +// division of a structure [k takes 2 bytes][m takes 2 bytes], then * 103 / 1024 +// will be [k / 10][m / 10]. It allows parallel division. +constexpr uint64_t kDivisionBy10Mul = 103u; constexpr uint64_t kDivisionBy10Div = 1 << 10; -// * kDivisionBy100Mul / kDivisionBy100Div is a division by 100 for values from -// 0 to 9999. -constexpr uint64_t kDivisionBy100Mul = - ComputePowerOf100DivisionCoefficient(2); -static_assert(kDivisionBy100Mul == 10486, - "division coefficient for 100 is incorrect"); +// * 10486 / 1048576 is a division by 100 for values from 0 to 9999. +constexpr uint64_t kDivisionBy100Mul = 10486u; constexpr uint64_t kDivisionBy100Div = 1 << 20; -static_assert(ComputePowerOf100DivisionCoefficient(3) == 1073742, - "division coefficient for 1000 is incorrect"); - -// Same as `PrepareEightDigits`, but produces 2 digits for integers < 100. -inline uint32_t PrepareTwoDigitsImpl(uint32_t i, bool reversed) { - assert(i < 100); - uint32_t div10 = (i * kDivisionBy10Mul) / kDivisionBy10Div; - uint32_t mod10 = i - 10u * div10; - return (div10 << (reversed ? 8 : 0)) + (mod10 << (reversed ? 0 : 8)); -} -inline uint32_t PrepareTwoDigits(uint32_t i) { - return PrepareTwoDigitsImpl(i, false); +// Encode functions write the ASCII output of input `n` to `out_str`. +inline char* EncodeHundred(uint32_t n, absl::Nonnull out_str) { + int num_digits = static_cast(n - 10) >> 8; + uint32_t div10 = (n * kDivisionBy10Mul) / kDivisionBy10Div; + uint32_t mod10 = n - 10u * div10; + uint32_t base = kTwoZeroBytes + div10 + (mod10 << 8); + base >>= num_digits & 8; + little_endian::Store16(out_str, static_cast(base)); + return out_str + 2 + num_digits; } -// Same as `PrepareEightDigits`, but produces 4 digits for integers < 10000. -inline uint32_t PrepareFourDigitsImpl(uint32_t n, bool reversed) { +inline char* EncodeTenThousand(uint32_t n, absl::Nonnull out_str) { // We split lower 2 digits and upper 2 digits of n into 2 byte consecutive // blocks. 123 -> [\0\1][\0\23]. We divide by 10 both blocks // (it's 1 division + zeroing upper bits), and compute modulo 10 as well "in @@ -231,19 +185,22 @@ inline uint32_t PrepareFourDigitsImpl(uint32_t n, bool reversed) { // strip trailing zeros, add ASCII '0000' and return. uint32_t div100 = (n * kDivisionBy100Mul) / kDivisionBy100Div; uint32_t mod100 = n - 100ull * div100; - uint32_t hundreds = - (mod100 << (reversed ? 0 : 16)) + (div100 << (reversed ? 16 : 0)); + uint32_t hundreds = (mod100 << 16) + div100; uint32_t tens = (hundreds * kDivisionBy10Mul) / kDivisionBy10Div; tens &= (0xFull << 16) | 0xFull; - tens = (tens << (reversed ? 8 : 0)) + - static_cast((hundreds - 10ull * tens) << (reversed ? 0 : 8)); - return tens; -} -inline uint32_t PrepareFourDigits(uint32_t n) { - return PrepareFourDigitsImpl(n, false); -} -inline uint32_t PrepareFourDigitsReversed(uint32_t n) { - return PrepareFourDigitsImpl(n, true); + tens += (hundreds - 10ull * tens) << 8; + ABSL_ASSUME(tens != 0); + // The result can contain trailing zero bits, we need to strip them to a first + // significant byte in a final representation. For example, for n = 123, we + // have tens to have representation \0\1\2\3. We do `& -8` to round + // to a multiple to 8 to strip zero bytes, not all zero bits. + // countr_zero to help. + // 0 minus 8 to make MSVC happy. + uint32_t zeroes = static_cast(absl::countr_zero(tens)) & (0 - 8u); + tens += kFourZeroBytes; + tens >>= zeroes; + little_endian::Store32(out_str, tens); + return out_str + sizeof(tens) - zeroes / 8; } // Helper function to produce an ASCII representation of `i`. @@ -259,309 +216,126 @@ inline uint32_t PrepareFourDigitsReversed(uint32_t n) { // // Note two leading zeros: // EXPECT_EQ(absl::string_view(ascii, 8), "00102030"); // -// If `Reversed` is set to true, the result becomes reversed to "03020100". -// // Pre-condition: `i` must be less than 100000000. -inline uint64_t PrepareEightDigitsImpl(uint32_t i, bool reversed) { +inline uint64_t PrepareEightDigits(uint32_t i) { ABSL_ASSUME(i < 10000'0000); // Prepare 2 blocks of 4 digits "in parallel". uint32_t hi = i / 10000; uint32_t lo = i % 10000; - uint64_t merged = (uint64_t{hi} << (reversed ? 32 : 0)) | - (uint64_t{lo} << (reversed ? 0 : 32)); + uint64_t merged = hi | (uint64_t{lo} << 32); uint64_t div100 = ((merged * kDivisionBy100Mul) / kDivisionBy100Div) & ((0x7Full << 32) | 0x7Full); uint64_t mod100 = merged - 100ull * div100; - uint64_t hundreds = - (mod100 << (reversed ? 0 : 16)) + (div100 << (reversed ? 16 : 0)); + uint64_t hundreds = (mod100 << 16) + div100; uint64_t tens = (hundreds * kDivisionBy10Mul) / kDivisionBy10Div; tens &= (0xFull << 48) | (0xFull << 32) | (0xFull << 16) | 0xFull; - tens = (tens << (reversed ? 8 : 0)) + - ((hundreds - 10ull * tens) << (reversed ? 0 : 8)); + tens += (hundreds - 10ull * tens) << 8; return tens; } -inline uint64_t PrepareEightDigits(uint32_t i) { - return PrepareEightDigitsImpl(i, false); -} -inline uint64_t PrepareEightDigitsReversed(uint32_t i) { - return PrepareEightDigitsImpl(i, true); -} -template -class FastUIntToStringConverter { - static_assert( - std::is_same())>::value, - "to avoid code bloat, only instantiate this for int and larger types"); - static_assert(std::is_unsigned::value, - "this class is only for unsigned types"); - - public: - // Outputs the given number backward (like with std::copy_backward), - // starting from the end of the string. - // The number of digits in the number must have been already measured and - // passed *exactly*, otherwise the behavior is undefined. - // (This is an optimization, as calculating the number of digits again would - // slow down the hot path.) - // Returns an iterator to the start of the suffix that was appended. - static BackwardIt FastIntToBufferBackward(T v, BackwardIt end) { - // THIS IS A HOT FUNCTION with a very deliberate structure to exploit branch - // prediction and shorten the critical path for smaller numbers. - // Do not move around the if/else blocks or attempt to simplify it - // without benchmarking any changes. - - if (v < 10) { - goto AT_LEAST_1 /* NOTE: mandatory for the 0 case */; - } - if (v < 1000) { - goto AT_LEAST_10; - } - if (v < 10000000) { - goto AT_LEAST_1000; - } - - if (v >= 100000000 / 10) { - if (v >= 10000000000000000 / 10) { - DoFastIntToBufferBackward<8>(v, end); - } - DoFastIntToBufferBackward<8>(v, end); - } - - if (v >= 10000 / 10) { - AT_LEAST_1000: - DoFastIntToBufferBackward<4>(v, end); - } - - if (v >= 100 / 10) { - AT_LEAST_10: - DoFastIntToBufferBackward<2>(v, end); - } - - if (v >= 10 / 10) { - AT_LEAST_1: - end = DoFastIntToBufferBackward(v, end, std::integral_constant()); - } - return end; +inline ABSL_ATTRIBUTE_ALWAYS_INLINE absl::Nonnull EncodeFullU32( + uint32_t n, absl::Nonnull out_str) { + if (n < 10) { + *out_str = static_cast('0' + n); + return out_str + 1; } - - private: - // Only assume pointers are contiguous for now. String and vector iterators - // could be special-cased as well, but there's no need for them here. - // With C++20 we can probably switch to std::contiguous_iterator_tag. - static constexpr bool kIsContiguousIterator = - std::is_pointer::value; - - template - static void DoFastIntToBufferBackward(T& v, BackwardIt& end) { - constexpr T kModulus = Pow(10, Exponent); - T remainder = static_cast(v % kModulus); - v = static_cast(v / kModulus); - end = DoFastIntToBufferBackward(remainder, end, - std::integral_constant()); - } - - static BackwardIt DoFastIntToBufferBackward(const T&, BackwardIt end, - std::integral_constant) { - return end; - } - - static BackwardIt DoFastIntToBufferBackward(T v, BackwardIt end, - std::integral_constant) { - *--end = static_cast('0' + v); - return DoFastIntToBufferBackward(v, end, std::integral_constant()); - } - - static BackwardIt DoFastIntToBufferBackward(T v, BackwardIt end, - std::integral_constant) { - if (kIsContiguousIterator) { - const uint32_t digits = - PrepareFourDigits(static_cast(v)) + kFourZeroBytes; - end -= sizeof(digits); - little_endian::Store32(&*end, digits); - } else { - uint32_t digits = - PrepareFourDigitsReversed(static_cast(v)) + kFourZeroBytes; - for (size_t i = 0; i < sizeof(digits); ++i) { - *--end = static_cast(digits); - digits >>= CHAR_BIT; - } - } - return end; + if (n < 100'000'000) { + uint64_t bottom = PrepareEightDigits(n); + ABSL_ASSUME(bottom != 0); + // 0 minus 8 to make MSVC happy. + uint32_t zeroes = + static_cast(absl::countr_zero(bottom)) & (0 - 8u); + little_endian::Store64(out_str, (bottom + kEightZeroBytes) >> zeroes); + return out_str + sizeof(bottom) - zeroes / 8; } - - static BackwardIt DoFastIntToBufferBackward(T v, BackwardIt end, - std::integral_constant) { - if (kIsContiguousIterator) { - const uint64_t digits = - PrepareEightDigits(static_cast(v)) + kEightZeroBytes; - end -= sizeof(digits); - little_endian::Store64(&*end, digits); - } else { - uint64_t digits = PrepareEightDigitsReversed(static_cast(v)) + - kEightZeroBytes; - for (size_t i = 0; i < sizeof(digits); ++i) { - *--end = static_cast(digits); - digits >>= CHAR_BIT; - } - } - return end; - } - - template - static BackwardIt DoFastIntToBufferBackward( - T v, BackwardIt end, std::integral_constant) { - constexpr int kLogModulus = Digits - Digits / 2; - constexpr T kModulus = Pow(static_cast(10), kLogModulus); - bool is_safe_to_use_division_trick = Digits <= 8; - T quotient, remainder; - if (is_safe_to_use_division_trick) { - constexpr uint64_t kCoefficient = - ComputePowerOf100DivisionCoefficient(kLogModulus); - quotient = (v * kCoefficient) >> (10 * kLogModulus); - remainder = v - quotient * kModulus; - } else { - quotient = v / kModulus; - remainder = v % kModulus; - } - end = DoFastIntToBufferBackward(remainder, end, - std::integral_constant()); - return DoFastIntToBufferBackward( - quotient, end, std::integral_constant()); - } -}; - -// Returns an iterator to the start of the suffix that was appended -template -std::enable_if_t::value, BackwardIt> -DoFastIntToBufferBackward(T v, BackwardIt end, uint32_t digits) { - using PromotedT = std::decay_t; - using Converter = FastUIntToStringConverter; - (void)digits; - return Converter().FastIntToBufferBackward(v, end); + uint32_t div08 = n / 100'000'000; + uint32_t mod08 = n % 100'000'000; + uint64_t bottom = PrepareEightDigits(mod08) + kEightZeroBytes; + out_str = EncodeHundred(div08, out_str); + little_endian::Store64(out_str, bottom); + return out_str + sizeof(bottom); } -template -std::enable_if_t::value, BackwardIt> -DoFastIntToBufferBackward(T v, BackwardIt end, uint32_t digits) { - if (absl::numbers_internal::IsNegative(v)) { - // Store the minus sign *before* we produce the number itself, not after. - // This gets us a tail call. - end[-static_cast(digits) - 1] = '-'; +inline ABSL_ATTRIBUTE_ALWAYS_INLINE char* EncodeFullU64(uint64_t i, + char* buffer) { + if (i <= std::numeric_limits::max()) { + return EncodeFullU32(static_cast(i), buffer); } - return DoFastIntToBufferBackward( - absl::numbers_internal::UnsignedAbsoluteValue(v), end, digits); -} - -template -std::enable_if_t::value, int> -GetNumDigitsOrNegativeIfNegativeImpl(T v) { - const auto /* either bool or std::false_type */ is_negative = - absl::numbers_internal::IsNegative(v); - const int digits = static_cast(absl::numbers_internal::Base10Digits( - absl::numbers_internal::UnsignedAbsoluteValue(v))); - return is_negative ? ~digits : digits; + uint32_t mod08; + if (i < 1'0000'0000'0000'0000ull) { + uint32_t div08 = static_cast(i / 100'000'000ull); + mod08 = static_cast(i % 100'000'000ull); + buffer = EncodeFullU32(div08, buffer); + } else { + uint64_t div08 = i / 100'000'000ull; + mod08 = static_cast(i % 100'000'000ull); + uint32_t div016 = static_cast(div08 / 100'000'000ull); + uint32_t div08mod08 = static_cast(div08 % 100'000'000ull); + uint64_t mid_result = PrepareEightDigits(div08mod08) + kEightZeroBytes; + buffer = EncodeTenThousand(div016, buffer); + little_endian::Store64(buffer, mid_result); + buffer += sizeof(mid_result); + } + uint64_t mod_result = PrepareEightDigits(mod08) + kEightZeroBytes; + little_endian::Store64(buffer, mod_result); + return buffer + sizeof(mod_result); } } // namespace void numbers_internal::PutTwoDigits(uint32_t i, absl::Nonnull buf) { - little_endian::Store16( - buf, static_cast(PrepareTwoDigits(i) + kTwoZeroBytes)); + assert(i < 100); + uint32_t base = kTwoZeroBytes; + uint32_t div10 = (i * kDivisionBy10Mul) / kDivisionBy10Div; + uint32_t mod10 = i - 10u * div10; + base += div10 + (mod10 << 8); + little_endian::Store16(buf, static_cast(base)); } absl::Nonnull numbers_internal::FastIntToBuffer( - uint32_t i, absl::Nonnull buffer) { - const uint32_t digits = absl::numbers_internal::Base10Digits(i); - buffer += digits; - *buffer = '\0'; // We're going backward, so store this first - FastIntToBufferBackward(i, buffer, digits); - return buffer; + uint32_t n, absl::Nonnull out_str) { + out_str = EncodeFullU32(n, out_str); + *out_str = '\0'; + return out_str; } absl::Nonnull numbers_internal::FastIntToBuffer( int32_t i, absl::Nonnull buffer) { - buffer += static_cast(i < 0); - uint32_t digits = absl::numbers_internal::Base10Digits( - absl::numbers_internal::UnsignedAbsoluteValue(i)); - buffer += digits; - *buffer = '\0'; // We're going backward, so store this first - FastIntToBufferBackward(i, buffer, digits); + uint32_t u = static_cast(i); + if (i < 0) { + *buffer++ = '-'; + // We need to do the negation in modular (i.e., "unsigned") + // arithmetic; MSVC++ apparently warns for plain "-u", so + // we write the equivalent expression "0 - u" instead. + u = 0 - u; + } + buffer = EncodeFullU32(u, buffer); + *buffer = '\0'; return buffer; } absl::Nonnull numbers_internal::FastIntToBuffer( uint64_t i, absl::Nonnull buffer) { - uint32_t digits = absl::numbers_internal::Base10Digits(i); - buffer += digits; - *buffer = '\0'; // We're going backward, so store this first - FastIntToBufferBackward(i, buffer, digits); + buffer = EncodeFullU64(i, buffer); + *buffer = '\0'; return buffer; } absl::Nonnull numbers_internal::FastIntToBuffer( int64_t i, absl::Nonnull buffer) { - buffer += static_cast(i < 0); - uint32_t digits = absl::numbers_internal::Base10Digits( - absl::numbers_internal::UnsignedAbsoluteValue(i)); - buffer += digits; - *buffer = '\0'; // We're going backward, so store this first - FastIntToBufferBackward(i, buffer, digits); + uint64_t u = static_cast(i); + if (i < 0) { + *buffer++ = '-'; + // We need to do the negation in modular (i.e., "unsigned") + // arithmetic; MSVC++ apparently warns for plain "-u", so + // we write the equivalent expression "0 - u" instead. + u = 0 - u; + } + buffer = EncodeFullU64(u, buffer); + *buffer = '\0'; return buffer; } -absl::Nonnull numbers_internal::FastIntToBufferBackward( - uint32_t i, absl::Nonnull buffer_end, uint32_t exact_digit_count) { - return DoFastIntToBufferBackward(i, buffer_end, exact_digit_count); -} - -absl::Nonnull numbers_internal::FastIntToBufferBackward( - int32_t i, absl::Nonnull buffer_end, uint32_t exact_digit_count) { - return DoFastIntToBufferBackward(i, buffer_end, exact_digit_count); -} - -absl::Nonnull numbers_internal::FastIntToBufferBackward( - uint64_t i, absl::Nonnull buffer_end, uint32_t exact_digit_count) { - return DoFastIntToBufferBackward(i, buffer_end, exact_digit_count); -} - -absl::Nonnull numbers_internal::FastIntToBufferBackward( - int64_t i, absl::Nonnull buffer_end, uint32_t exact_digit_count) { - return DoFastIntToBufferBackward(i, buffer_end, exact_digit_count); -} - -int numbers_internal::GetNumDigitsOrNegativeIfNegative(signed char v) { - return GetNumDigitsOrNegativeIfNegativeImpl(v); -} -int numbers_internal::GetNumDigitsOrNegativeIfNegative(unsigned char v) { - return GetNumDigitsOrNegativeIfNegativeImpl(v); -} -int numbers_internal::GetNumDigitsOrNegativeIfNegative(short v) { // NOLINT - return GetNumDigitsOrNegativeIfNegativeImpl(v); -} -int numbers_internal::GetNumDigitsOrNegativeIfNegative( - unsigned short v) { // NOLINT - return GetNumDigitsOrNegativeIfNegativeImpl(v); -} -int numbers_internal::GetNumDigitsOrNegativeIfNegative(int v) { - return GetNumDigitsOrNegativeIfNegativeImpl(v); -} -int numbers_internal::GetNumDigitsOrNegativeIfNegative(unsigned int v) { - return GetNumDigitsOrNegativeIfNegativeImpl(v); -} -int numbers_internal::GetNumDigitsOrNegativeIfNegative(long v) { // NOLINT - return GetNumDigitsOrNegativeIfNegativeImpl(v); -} -int numbers_internal::GetNumDigitsOrNegativeIfNegative( - unsigned long v) { // NOLINT - return GetNumDigitsOrNegativeIfNegativeImpl(v); -} -int numbers_internal::GetNumDigitsOrNegativeIfNegative(long long v) { // NOLINT - return GetNumDigitsOrNegativeIfNegativeImpl(v); -} -int numbers_internal::GetNumDigitsOrNegativeIfNegative( - unsigned long long v) { // NOLINT - return GetNumDigitsOrNegativeIfNegativeImpl(v); -} - // Given a 128-bit number expressed as a pair of uint64_t, high half first, // return that number multiplied by the given 32-bit value. If the result is // too large to fit in a 128-bit number, divide it by 2 until it fits. diff --git a/absl/strings/numbers.h b/absl/strings/numbers.h index ad4e66b6..739dbb28 100644 --- a/absl/strings/numbers.h +++ b/absl/strings/numbers.h @@ -32,7 +32,6 @@ #endif #include -#include #include #include #include @@ -40,12 +39,10 @@ #include #include -#include "absl/base/attributes.h" #include "absl/base/config.h" #include "absl/base/internal/endian.h" #include "absl/base/macros.h" #include "absl/base/nullability.h" -#include "absl/base/optimization.h" #include "absl/base/port.h" #include "absl/numeric/bits.h" #include "absl/numeric/int128.h" @@ -161,96 +158,6 @@ bool safe_strtou128_base(absl::string_view text, static const int kFastToBufferSize = 32; static const int kSixDigitsToBufferSize = 16; -template -std::enable_if_t::value, bool> IsNegative(const T& v) { - return v < T(); -} - -template -std::enable_if_t::value, std::false_type> IsNegative( - const T&) { - // The integer is unsigned, so return a compile-time constant. - // This can help the optimizer avoid having to prove bool to be false later. - return std::false_type(); -} - -template -std::enable_if_t>::value, T&&> -UnsignedAbsoluteValue(T&& v ABSL_ATTRIBUTE_LIFETIME_BOUND) { - // The value is unsigned; just return the original. - return std::forward(v); -} - -template -ABSL_ATTRIBUTE_CONST_FUNCTION - std::enable_if_t::value, std::make_unsigned_t> - UnsignedAbsoluteValue(T v) { - using U = std::make_unsigned_t; - return IsNegative(v) ? U() - static_cast(v) : static_cast(v); -} - -// Returns the number of base-10 digits in the given number. -// Note that this strictly counts digits. It does not count the sign. -// The `initial_digits` parameter is the starting point, which is normally equal -// to 1 because the number of digits in 0 is 1 (a special case). -// However, callers may e.g. wish to change it to 2 to account for the sign. -template -std::enable_if_t::value, uint32_t> Base10Digits( - T v, const uint32_t initial_digits = 1) { - uint32_t r = initial_digits; - // If code size becomes an issue, the 'if' stage can be removed for a minor - // performance loss. - for (;;) { - if (ABSL_PREDICT_TRUE(v < 10 * 10)) { - r += (v >= 10); - break; - } - if (ABSL_PREDICT_TRUE(v < 1000 * 10)) { - r += (v >= 1000) + 2; - break; - } - if (ABSL_PREDICT_TRUE(v < 100000 * 10)) { - r += (v >= 100000) + 4; - break; - } - r += 6; - v = static_cast(v / 1000000); - } - return r; -} - -template -std::enable_if_t::value, uint32_t> Base10Digits( - T v, uint32_t r = 1) { - // Branchlessly add 1 to account for a minus sign. - r += static_cast(IsNegative(v)); - return Base10Digits(UnsignedAbsoluteValue(v), r); -} - -// These functions return the number of base-10 digits, but multiplied by -1 if -// the input itself is negative. This is handy and efficient for later usage, -// since the bitwise complement of the result becomes equal to the number of -// characters required. -ABSL_ATTRIBUTE_CONST_FUNCTION int GetNumDigitsOrNegativeIfNegative( - signed char v); -ABSL_ATTRIBUTE_CONST_FUNCTION int GetNumDigitsOrNegativeIfNegative( - unsigned char v); -ABSL_ATTRIBUTE_CONST_FUNCTION int GetNumDigitsOrNegativeIfNegative( - short v); // NOLINT -ABSL_ATTRIBUTE_CONST_FUNCTION int GetNumDigitsOrNegativeIfNegative( - unsigned short v); // NOLINT -ABSL_ATTRIBUTE_CONST_FUNCTION int GetNumDigitsOrNegativeIfNegative(int v); -ABSL_ATTRIBUTE_CONST_FUNCTION int GetNumDigitsOrNegativeIfNegative( - unsigned int v); -ABSL_ATTRIBUTE_CONST_FUNCTION int GetNumDigitsOrNegativeIfNegative( - long v); // NOLINT -ABSL_ATTRIBUTE_CONST_FUNCTION int GetNumDigitsOrNegativeIfNegative( - unsigned long v); // NOLINT -ABSL_ATTRIBUTE_CONST_FUNCTION int GetNumDigitsOrNegativeIfNegative( - long long v); // NOLINT -ABSL_ATTRIBUTE_CONST_FUNCTION int GetNumDigitsOrNegativeIfNegative( - unsigned long long v); // NOLINT - // Helper function for fast formatting of floating-point values. // The result is the same as printf's "%g", a.k.a. "%.6g"; that is, six // significant digits are returned, trailing zeros are removed, and numbers @@ -259,18 +166,24 @@ ABSL_ATTRIBUTE_CONST_FUNCTION int GetNumDigitsOrNegativeIfNegative( // Required buffer size is `kSixDigitsToBufferSize`. size_t SixDigitsToBuffer(double d, absl::Nonnull buffer); -// All of these functions take an output buffer +// WARNING: These functions may write more characters than necessary, because +// they are intended for speed. All functions take an output buffer // as an argument and return a pointer to the last byte they wrote, which is the // terminating '\0'. At most `kFastToBufferSize` bytes are written. -absl::Nonnull FastIntToBuffer(int32_t i, absl::Nonnull buffer); -absl::Nonnull FastIntToBuffer(uint32_t i, absl::Nonnull buffer); -absl::Nonnull FastIntToBuffer(int64_t i, absl::Nonnull buffer); -absl::Nonnull FastIntToBuffer(uint64_t i, absl::Nonnull buffer); +absl::Nonnull FastIntToBuffer(int32_t i, absl::Nonnull buffer) + ABSL_INTERNAL_NEED_MIN_SIZE(buffer, kFastToBufferSize); +absl::Nonnull FastIntToBuffer(uint32_t n, absl::Nonnull out_str) + ABSL_INTERNAL_NEED_MIN_SIZE(out_str, kFastToBufferSize); +absl::Nonnull FastIntToBuffer(int64_t i, absl::Nonnull buffer) + ABSL_INTERNAL_NEED_MIN_SIZE(buffer, kFastToBufferSize); +absl::Nonnull FastIntToBuffer(uint64_t i, absl::Nonnull buffer) + ABSL_INTERNAL_NEED_MIN_SIZE(buffer, kFastToBufferSize); // For enums and integer types that are not an exact match for the types above, // use templates to call the appropriate one of the four overloads above. template -absl::Nonnull FastIntToBuffer(int_type i, absl::Nonnull buffer) { +absl::Nonnull FastIntToBuffer(int_type i, absl::Nonnull buffer) + ABSL_INTERNAL_NEED_MIN_SIZE(buffer, kFastToBufferSize) { static_assert(sizeof(i) <= 64 / 8, "FastIntToBuffer works only with 64-bit-or-less integers."); // TODO(jorg): This signed-ness check is used because it works correctly @@ -294,58 +207,6 @@ absl::Nonnull FastIntToBuffer(int_type i, absl::Nonnull buffer) { } } -// These functions do NOT add any null-terminator. -// They return a pointer to the beginning of the written string. -// The digit counts provided must *exactly* match the number of base-10 digits -// in the number, or the behavior is undefined. -// (i.e. do NOT count the minus sign, or over- or under-count the digits.) -absl::Nonnull FastIntToBufferBackward(int32_t i, - absl::Nonnull buffer_end, - uint32_t exact_digit_count); -absl::Nonnull FastIntToBufferBackward(uint32_t i, - absl::Nonnull buffer_end, - uint32_t exact_digit_count); -absl::Nonnull FastIntToBufferBackward(int64_t i, - absl::Nonnull buffer_end, - uint32_t exact_digit_count); -absl::Nonnull FastIntToBufferBackward(uint64_t i, - absl::Nonnull buffer_end, - uint32_t exact_digit_count); - -// For enums and integer types that are not an exact match for the types above, -// use templates to call the appropriate one of the four overloads above. -template -absl::Nonnull FastIntToBufferBackward(int_type i, - absl::Nonnull buffer_end, - uint32_t exact_digit_count) { - static_assert( - sizeof(i) <= 64 / 8, - "FastIntToBufferBackward works only with 64-bit-or-less integers."); - // This signed-ness check is used because it works correctly - // with enums, and it also serves to check that int_type is not a pointer. - // If one day something like std::is_signed works, switch to it. - // These conditions are constexpr bools to suppress MSVC warning C4127. - constexpr bool kIsSigned = static_cast(1) - 2 < 0; - constexpr bool kUse64Bit = sizeof(i) > 32 / 8; - if (kIsSigned) { - if (kUse64Bit) { - return FastIntToBufferBackward(static_cast(i), buffer_end, - exact_digit_count); - } else { - return FastIntToBufferBackward(static_cast(i), buffer_end, - exact_digit_count); - } - } else { - if (kUse64Bit) { - return FastIntToBufferBackward(static_cast(i), buffer_end, - exact_digit_count); - } else { - return FastIntToBufferBackward(static_cast(i), buffer_end, - exact_digit_count); - } - } -} - // Implementation of SimpleAtoi, generalized to support arbitrary base (used // with base different from 10 elsewhere in Abseil implementation). template diff --git a/absl/strings/numbers_test.cc b/absl/strings/numbers_test.cc index 1ceff70f..75c2dcf2 100644 --- a/absl/strings/numbers_test.cc +++ b/absl/strings/numbers_test.cc @@ -231,15 +231,10 @@ TEST(Numbers, TestFastPrints) { CheckInt32(INT_MIN); CheckInt32(INT_MAX); CheckInt64(LONG_MIN); - CheckInt64(uint64_t{10000000}); - CheckInt64(uint64_t{100000000}); CheckInt64(uint64_t{1000000000}); CheckInt64(uint64_t{9999999999}); CheckInt64(uint64_t{100000000000000}); CheckInt64(uint64_t{999999999999999}); - CheckInt64(uint64_t{1000000000000000}); - CheckInt64(uint64_t{10000000000000000}); - CheckInt64(uint64_t{100000000000000000}); CheckInt64(uint64_t{1000000000000000000}); CheckInt64(uint64_t{1199999999999999999}); CheckInt64(int64_t{-700000000000000000}); @@ -251,8 +246,6 @@ TEST(Numbers, TestFastPrints) { CheckUInt64(uint64_t{999999999999999}); CheckUInt64(uint64_t{1000000000000000000}); CheckUInt64(uint64_t{1199999999999999999}); - CheckUInt64(uint64_t{10000000000000000000u}); - CheckUInt64(uint64_t{10200300040000500006u}); CheckUInt64(std::numeric_limits::max()); for (int i = 0; i < 10000; i++) { diff --git a/absl/strings/str_cat.cc b/absl/strings/str_cat.cc index 098ab183..e7f7052a 100644 --- a/absl/strings/str_cat.cc +++ b/absl/strings/str_cat.cc @@ -21,12 +21,10 @@ #include #include #include -#include #include "absl/base/config.h" #include "absl/base/nullability.h" #include "absl/strings/internal/resize_uninitialized.h" -#include "absl/strings/numbers.h" #include "absl/strings/string_view.h" namespace absl { @@ -43,7 +41,8 @@ ABSL_NAMESPACE_BEGIN namespace { // Append is merely a version of memcpy that returns the address of the byte // after the area just overwritten. -absl::Nonnull Append(absl::Nonnull out, const AlphaNum& x) { +inline absl::Nonnull Append(absl::Nonnull out, + const AlphaNum& x) { // memcpy is allowed to overwrite arbitrary memory, so doing this after the // call would force an extra fetch of x.size(). char* after = out + x.size(); @@ -53,6 +52,11 @@ absl::Nonnull Append(absl::Nonnull out, const AlphaNum& x) { return after; } +inline void STLStringAppendUninitializedAmortized(std::string* dest, + size_t to_append) { + strings_internal::AppendUninitializedTraits::Append(dest, + to_append); +} } // namespace std::string StrCat(const AlphaNum& a, const AlphaNum& b) { @@ -98,130 +102,6 @@ std::string StrCat(const AlphaNum& a, const AlphaNum& b, const AlphaNum& c, namespace strings_internal { // Do not call directly - these are not part of the public API. -void STLStringAppendUninitializedAmortized(std::string* dest, - size_t to_append) { - strings_internal::AppendUninitializedTraits::Append(dest, - to_append); -} - -template -std::enable_if_t::value, std::string> IntegerToString( - Integer i) { - std::string str; - const auto /* either bool or std::false_type */ is_negative = - absl::numbers_internal::IsNegative(i); - const uint32_t digits = absl::numbers_internal::Base10Digits( - absl::numbers_internal::UnsignedAbsoluteValue(i)); - absl::strings_internal::STLStringResizeUninitialized( - &str, digits + static_cast(is_negative)); - absl::numbers_internal::FastIntToBufferBackward(i, &str[str.size()], digits); - return str; -} - -template <> -std::string IntegerToString(long i) { // NOLINT - if (sizeof(i) <= sizeof(int)) { - return IntegerToString(static_cast(i)); - } else { - return IntegerToString(static_cast(i)); // NOLINT - } -} - -template <> -std::string IntegerToString(unsigned long i) { // NOLINT - if (sizeof(i) <= sizeof(unsigned int)) { - return IntegerToString(static_cast(i)); - } else { - return IntegerToString(static_cast(i)); // NOLINT - } -} - -template -std::enable_if_t::value, std::string> -FloatToString(Float f) { - std::string result; - strings_internal::STLStringResizeUninitialized( - &result, numbers_internal::kSixDigitsToBufferSize); - char* start = &result[0]; - result.erase(numbers_internal::SixDigitsToBuffer(f, start)); - return result; -} - -std::string SingleArgStrCat(int x) { return IntegerToString(x); } -std::string SingleArgStrCat(unsigned int x) { return IntegerToString(x); } -// NOLINTNEXTLINE -std::string SingleArgStrCat(long x) { return IntegerToString(x); } -// NOLINTNEXTLINE -std::string SingleArgStrCat(unsigned long x) { return IntegerToString(x); } -// NOLINTNEXTLINE -std::string SingleArgStrCat(long long x) { return IntegerToString(x); } -// NOLINTNEXTLINE -std::string SingleArgStrCat(unsigned long long x) { return IntegerToString(x); } -std::string SingleArgStrCat(float x) { return FloatToString(x); } -std::string SingleArgStrCat(double x) { return FloatToString(x); } - -template -std::enable_if_t::value, void> AppendIntegerToString( - std::string& str, Integer i) { - const auto /* either bool or std::false_type */ is_negative = - absl::numbers_internal::IsNegative(i); - const uint32_t digits = absl::numbers_internal::Base10Digits( - absl::numbers_internal::UnsignedAbsoluteValue(i)); - absl::strings_internal::STLStringAppendUninitializedAmortized( - &str, digits + static_cast(is_negative)); - absl::numbers_internal::FastIntToBufferBackward(i, &str[str.size()], digits); -} - -template <> -void AppendIntegerToString(std::string& str, long i) { // NOLINT - if (sizeof(i) <= sizeof(int)) { - return AppendIntegerToString(str, static_cast(i)); - } else { - return AppendIntegerToString(str, static_cast(i)); // NOLINT - } -} - -template <> -void AppendIntegerToString(std::string& str, - unsigned long i) { // NOLINT - if (sizeof(i) <= sizeof(unsigned int)) { - return AppendIntegerToString(str, static_cast(i)); - } else { - return AppendIntegerToString(str, - static_cast(i)); // NOLINT - } -} - -// `SingleArgStrAppend` overloads are defined here for the same reasons as with -// `SingleArgStrCat` above. -void SingleArgStrAppend(std::string& str, int x) { - return AppendIntegerToString(str, x); -} - -void SingleArgStrAppend(std::string& str, unsigned int x) { - return AppendIntegerToString(str, x); -} - -// NOLINTNEXTLINE -void SingleArgStrAppend(std::string& str, long x) { - return AppendIntegerToString(str, x); -} - -// NOLINTNEXTLINE -void SingleArgStrAppend(std::string& str, unsigned long x) { - return AppendIntegerToString(str, x); -} - -// NOLINTNEXTLINE -void SingleArgStrAppend(std::string& str, long long x) { - return AppendIntegerToString(str, x); -} - -// NOLINTNEXTLINE -void SingleArgStrAppend(std::string& str, unsigned long long x) { - return AppendIntegerToString(str, x); -} - std::string CatPieces(std::initializer_list pieces) { std::string result; size_t total_size = 0; @@ -258,7 +138,7 @@ void AppendPieces(absl::Nonnull dest, ASSERT_NO_OVERLAP(*dest, piece); to_append += piece.size(); } - strings_internal::STLStringAppendUninitializedAmortized(dest, to_append); + STLStringAppendUninitializedAmortized(dest, to_append); char* const begin = &(*dest)[0]; char* out = begin + old_size; @@ -277,7 +157,7 @@ void AppendPieces(absl::Nonnull dest, void StrAppend(absl::Nonnull dest, const AlphaNum& a) { ASSERT_NO_OVERLAP(*dest, a); std::string::size_type old_size = dest->size(); - strings_internal::STLStringAppendUninitializedAmortized(dest, a.size()); + STLStringAppendUninitializedAmortized(dest, a.size()); char* const begin = &(*dest)[0]; char* out = begin + old_size; out = Append(out, a); @@ -289,8 +169,7 @@ void StrAppend(absl::Nonnull dest, const AlphaNum& a, ASSERT_NO_OVERLAP(*dest, a); ASSERT_NO_OVERLAP(*dest, b); std::string::size_type old_size = dest->size(); - strings_internal::STLStringAppendUninitializedAmortized(dest, - a.size() + b.size()); + STLStringAppendUninitializedAmortized(dest, a.size() + b.size()); char* const begin = &(*dest)[0]; char* out = begin + old_size; out = Append(out, a); @@ -304,8 +183,7 @@ void StrAppend(absl::Nonnull dest, const AlphaNum& a, ASSERT_NO_OVERLAP(*dest, b); ASSERT_NO_OVERLAP(*dest, c); std::string::size_type old_size = dest->size(); - strings_internal::STLStringAppendUninitializedAmortized( - dest, a.size() + b.size() + c.size()); + STLStringAppendUninitializedAmortized(dest, a.size() + b.size() + c.size()); char* const begin = &(*dest)[0]; char* out = begin + old_size; out = Append(out, a); @@ -321,7 +199,7 @@ void StrAppend(absl::Nonnull dest, const AlphaNum& a, ASSERT_NO_OVERLAP(*dest, c); ASSERT_NO_OVERLAP(*dest, d); std::string::size_type old_size = dest->size(); - strings_internal::STLStringAppendUninitializedAmortized( + STLStringAppendUninitializedAmortized( dest, a.size() + b.size() + c.size() + d.size()); char* const begin = &(*dest)[0]; char* out = begin + old_size; diff --git a/absl/strings/str_cat.h b/absl/strings/str_cat.h index 1b52a36f..f224942f 100644 --- a/absl/strings/str_cat.h +++ b/absl/strings/str_cat.h @@ -93,6 +93,8 @@ #include #include #include +#include +#include #include #include #include @@ -448,36 +450,77 @@ std::string CatPieces(std::initializer_list pieces); void AppendPieces(absl::Nonnull dest, std::initializer_list pieces); -void STLStringAppendUninitializedAmortized(std::string* dest, size_t to_append); +template +std::string IntegerToString(Integer i) { + // Any integer (signed/unsigned) up to 64 bits can be formatted into a buffer + // with 22 bytes (including NULL at the end). + constexpr size_t kMaxDigits10 = 22; + std::string result; + strings_internal::STLStringResizeUninitialized(&result, kMaxDigits10); + char* start = &result[0]; + // note: this can be optimized to not write last zero. + char* end = numbers_internal::FastIntToBuffer(i, start); + auto size = static_cast(end - start); + assert((size < result.size()) && + "StrCat(Integer) does not fit into kMaxDigits10"); + result.erase(size); + return result; +} +template +std::string FloatToString(Float f) { + std::string result; + strings_internal::STLStringResizeUninitialized( + &result, numbers_internal::kSixDigitsToBufferSize); + char* start = &result[0]; + result.erase(numbers_internal::SixDigitsToBuffer(f, start)); + return result; +} // `SingleArgStrCat` overloads take built-in `int`, `long` and `long long` types // (signed / unsigned) to avoid ambiguity on the call side. If we used int32_t // and int64_t, then at least one of the three (`int` / `long` / `long long`) // would have been ambiguous when passed to `SingleArgStrCat`. -std::string SingleArgStrCat(int x); -std::string SingleArgStrCat(unsigned int x); -std::string SingleArgStrCat(long x); // NOLINT -std::string SingleArgStrCat(unsigned long x); // NOLINT -std::string SingleArgStrCat(long long x); // NOLINT -std::string SingleArgStrCat(unsigned long long x); // NOLINT -std::string SingleArgStrCat(float x); -std::string SingleArgStrCat(double x); - -// `SingleArgStrAppend` overloads are defined here for the same reasons as with -// `SingleArgStrCat` above. -void SingleArgStrAppend(std::string& str, int x); -void SingleArgStrAppend(std::string& str, unsigned int x); -void SingleArgStrAppend(std::string& str, long x); // NOLINT -void SingleArgStrAppend(std::string& str, unsigned long x); // NOLINT -void SingleArgStrAppend(std::string& str, long long x); // NOLINT -void SingleArgStrAppend(std::string& str, unsigned long long x); // NOLINT - -template ::value && - !std::is_same::value && - !std::is_same::value>> +inline std::string SingleArgStrCat(int x) { return IntegerToString(x); } +inline std::string SingleArgStrCat(unsigned int x) { + return IntegerToString(x); +} +// NOLINTNEXTLINE +inline std::string SingleArgStrCat(long x) { return IntegerToString(x); } +// NOLINTNEXTLINE +inline std::string SingleArgStrCat(unsigned long x) { + return IntegerToString(x); +} +// NOLINTNEXTLINE +inline std::string SingleArgStrCat(long long x) { return IntegerToString(x); } +// NOLINTNEXTLINE +inline std::string SingleArgStrCat(unsigned long long x) { + return IntegerToString(x); +} +inline std::string SingleArgStrCat(float x) { return FloatToString(x); } +inline std::string SingleArgStrCat(double x) { return FloatToString(x); } + +// As of September 2023, the SingleArgStrCat() optimization is only enabled for +// libc++. The reasons for this are: +// 1) The SSO size for libc++ is 23, while libstdc++ and MSSTL have an SSO size +// of 15. Since IntegerToString unconditionally resizes the string to 22 bytes, +// this causes both libstdc++ and MSSTL to allocate. +// 2) strings_internal::STLStringResizeUninitialized() only has an +// implementation that avoids initialization when using libc++. This isn't as +// relevant as (1), and the cost should be benchmarked if (1) ever changes on +// libstc++ or MSSTL. +#ifdef _LIBCPP_VERSION +#define ABSL_INTERNAL_STRCAT_ENABLE_FAST_CASE true +#else +#define ABSL_INTERNAL_STRCAT_ENABLE_FAST_CASE false +#endif + +template {} && !std::is_same{}>> using EnableIfFastCase = T; +#undef ABSL_INTERNAL_STRCAT_ENABLE_FAST_CASE + } // namespace strings_internal ABSL_MUST_USE_RESULT inline std::string StrCat() { return std::string(); } @@ -553,67 +596,6 @@ inline void StrAppend(absl::Nonnull dest, const AlphaNum& a, static_cast(args).Piece()...}); } -template -std::enable_if_t< - std::is_integral>::value, void> -StrAppend(absl::Nonnull result, T i) { - return absl::strings_internal::SingleArgStrAppend(*result, i); -} - -// This overload is only selected if all the parameters are numbers that can be -// handled quickly. -// Later we can look into how we can extend this to more general argument -// mixtures without bloating codegen too much, or copying unnecessarily. -template -std::enable_if_t< - (sizeof...(T) > 1), - std::common_type_t>...>> -StrAppend(absl::Nonnull str, T... args) { - // Do not add unnecessary variables, logic, or even "free" lambdas here. - // They can add overhead for the compiler and/or at run time. - // Furthermore, assume this function will be inlined. - // This function is carefully tailored to be able to be largely optimized away - // so that it becomes near-equivalent to the caller handling each argument - // individually while minimizing register pressure, so that the compiler - // can inline it with minimal overhead. - - // First, calculate the total length, so we can perform just a single resize. - // Save all the lengths for later. - size_t total_length = 0; - const ptrdiff_t lengths[] = { - absl::numbers_internal::GetNumDigitsOrNegativeIfNegative(args)...}; - for (const ptrdiff_t possibly_negative_length : lengths) { - // Lengths are negative for negative numbers. Keep them for later use, but - // take their absolute values for calculating total lengths; - total_length += possibly_negative_length < 0 - ? static_cast(-possibly_negative_length) - : static_cast(possibly_negative_length); - } - - // Now reserve space for all the arguments. - const size_t old_size = str->size(); - absl::strings_internal::STLStringAppendUninitializedAmortized(str, - total_length); - - // Finally, output each argument one-by-one, from left to right. - size_t i = 0; // The current argument we're processing - ptrdiff_t n; // The length of the current argument - typename String::pointer pos = &(*str)[old_size]; - using SomeTrivialEmptyType = std::false_type; - const SomeTrivialEmptyType dummy; - // Ugly code due to the lack of C++17 fold expressions - const SomeTrivialEmptyType dummies[] = { - (/* Comma expressions are poor man's C++17 fold expression for C++14 */ - (void)(n = lengths[i]), - (void)(n < 0 ? (void)(*pos++ = '-'), (n = ~n) : 0), - (void)absl::numbers_internal::FastIntToBufferBackward( - absl::numbers_internal::UnsignedAbsoluteValue(std::move(args)), - pos += n, static_cast(n)), - (void)++i, dummy)...}; - (void)dummies; // Remove & migrate to fold expressions in C++17 -} - // Helper function for the future StrCat default floating-point format, %.6g // This is fast. inline strings_internal::AlphaNumBuffer< diff --git a/absl/strings/str_cat_test.cc b/absl/strings/str_cat_test.cc index b30a86fe..66eddf0d 100644 --- a/absl/strings/str_cat_test.cc +++ b/absl/strings/str_cat_test.cc @@ -39,24 +39,6 @@ namespace { -template -void VerifyInteger(Integer value) { - const std::string expected = std::to_string(value); - - EXPECT_EQ(absl::StrCat(value), expected); - - const char* short_prefix = "x"; - const char* long_prefix = "2;k.msabxiuow2[09i;o3k21-93-9=29]"; - - std::string short_str = short_prefix; - absl::StrAppend(&short_str, value); - EXPECT_EQ(short_str, short_prefix + expected); - - std::string long_str = long_prefix; - absl::StrAppend(&long_str, value); - EXPECT_EQ(long_str, long_prefix + expected); -} - // Test absl::StrCat of ints and longs of various sizes and signdedness. TEST(StrCat, Ints) { const short s = -1; // NOLINT(runtime/int) @@ -86,34 +68,6 @@ TEST(StrCat, Ints) { EXPECT_EQ(answer, "-9-12"); answer = absl::StrCat(uintptr, 0); EXPECT_EQ(answer, "130"); - - for (const uint32_t base : {2u, 10u}) { - for (const int extra_shift : {0, 12}) { - for (uint64_t i = 0; i < (1 << 8); ++i) { - uint64_t j = i; - while (true) { - uint64_t v = j ^ (extra_shift != 0 ? (j << extra_shift) * base : 0); - VerifyInteger(static_cast(v)); - VerifyInteger(static_cast(v)); - VerifyInteger(static_cast(v)); - VerifyInteger(static_cast(v)); - VerifyInteger(static_cast(v)); // NOLINT - VerifyInteger(static_cast(v)); // NOLINT - VerifyInteger(static_cast(v)); // NOLINT - VerifyInteger(static_cast(v)); // NOLINT - VerifyInteger(static_cast(v)); // NOLINT - VerifyInteger(static_cast(v)); // NOLINT - VerifyInteger(static_cast(v)); // NOLINT - VerifyInteger(static_cast(v)); // NOLINT - const uint64_t next = j == 0 ? 1 : j * base; - if (next <= j) { - break; - } - j = next; - } - } - } - } } TEST(StrCat, Enums) { -- 2.34.1