Restore the simple loop in fastfloat_strncasecmp PR #356 replaced the 3/5-character case-insensitive compare used by parse_infnan with SWAR helpers (fastfloat_strncasecmp3/5 plus a generic memcpy-based fastfloat_strncasecmp). That code only runs for "nan", "inf" and "infinity" inputs, so it cannot help ordinary parsing, but it grew the inlined inf/nan tail of the parser: on x86-64 (GCC 14 and clang 21) the per-commit benchmarks dropped by 1.5-2.5% because the extra dead code changed inlining decisions and register allocation on the hot path (+7 instructions/float on canada.txt). Go back to the plain loop, which the compiler unrolls for the constant lengths 3 and 5, and drop the now-unused generic SWAR function. The `actual < 256` guard of the original is unnecessary: OR-ing 0x20 into a value can only equal an ASCII lowercase letter when the value is that letter or its uppercase form, so the plain OR is exact for every character type.
diff --git a/include/fast_float/float_common.h b/include/fast_float/float_common.h index 3769972..1b2594e 100644 --- a/include/fast_float/float_common.h +++ b/include/fast_float/float_common.h
@@ -325,147 +325,22 @@ > { }; -template <typename UC> -inline FASTFLOAT_CONSTEXPR14 bool -fastfloat_strncasecmp3(UC const *actual_mixedcase, - UC const *expected_lowercase) { - uint64_t mask{0}; - FASTFLOAT_IF_CONSTEXPR17(sizeof(UC) == 1) { mask = 0x2020202020202020; } - else FASTFLOAT_IF_CONSTEXPR17(sizeof(UC) == 2) { - mask = 0x0020002000200020; - } - else FASTFLOAT_IF_CONSTEXPR17(sizeof(UC) == 4) { - mask = 0x0000002000000020; - } - else { - return false; - } - - uint64_t val1{0}, val2{0}; - if (cpp20_and_in_constexpr()) { - for (size_t i = 0; i < 3; i++) { - if ((actual_mixedcase[i] | 32) != expected_lowercase[i]) { - return false; - } - } - return true; - } else { - FASTFLOAT_IF_CONSTEXPR17(sizeof(UC) == 1 || sizeof(UC) == 2) { - ::memcpy(&val1, actual_mixedcase, 3 * sizeof(UC)); - ::memcpy(&val2, expected_lowercase, 3 * sizeof(UC)); - val1 |= mask; - val2 |= mask; - return val1 == val2; - } - else FASTFLOAT_IF_CONSTEXPR17(sizeof(UC) == 4) { - ::memcpy(&val1, actual_mixedcase, 2 * sizeof(UC)); - ::memcpy(&val2, expected_lowercase, 2 * sizeof(UC)); - val1 |= mask; - if (val1 != val2) { - return false; - } - return (actual_mixedcase[2] | 32) == (expected_lowercase[2]); - } - else { - return false; - } - } -} - -template <typename UC> -inline FASTFLOAT_CONSTEXPR14 bool -fastfloat_strncasecmp5(UC const *actual_mixedcase, - UC const *expected_lowercase) { - uint64_t mask{0}; - uint64_t val1{0}, val2{0}; - if (cpp20_and_in_constexpr()) { - for (size_t i = 0; i < 5; i++) { - if ((actual_mixedcase[i] | 32) != expected_lowercase[i]) { - return false; - } - } - return true; - } else { - FASTFLOAT_IF_CONSTEXPR17(sizeof(UC) == 1) { - mask = 0x2020202020202020; - ::memcpy(&val1, actual_mixedcase, 5 * sizeof(UC)); - ::memcpy(&val2, expected_lowercase, 5 * sizeof(UC)); - val1 |= mask; - val2 |= mask; - return val1 == val2; - } - else FASTFLOAT_IF_CONSTEXPR17(sizeof(UC) == 2) { - mask = 0x0020002000200020; - ::memcpy(&val1, actual_mixedcase, 4 * sizeof(UC)); - ::memcpy(&val2, expected_lowercase, 4 * sizeof(UC)); - val1 |= mask; - if (val1 != val2) { - return false; - } - return (actual_mixedcase[4] | 32) == (expected_lowercase[4]); - } - else FASTFLOAT_IF_CONSTEXPR17(sizeof(UC) == 4) { - mask = 0x0000002000000020; - ::memcpy(&val1, actual_mixedcase, 2 * sizeof(UC)); - ::memcpy(&val2, expected_lowercase, 2 * sizeof(UC)); - val1 |= mask; - if (val1 != val2) { - return false; - } - ::memcpy(&val1, actual_mixedcase + 2, 2 * sizeof(UC)); - ::memcpy(&val2, expected_lowercase + 2, 2 * sizeof(UC)); - val1 |= mask; - if (val1 != val2) { - return false; - } - return (actual_mixedcase[4] | 32) == (expected_lowercase[4]); - } - else { - return false; - } - } -} - -// Compares two ASCII strings in a case insensitive manner. +// Compares two ASCII strings in a case insensitive manner. The expected +// string is lowercase ASCII, so OR-ing 0x20 into each actual character maps +// 'A'..'Z' onto 'a'..'z' and leaves everything else mismatched. The lengths +// used are tiny constants (3 and 5), so the compiler unrolls this loop; keep +// it small so that the (rarely executed) inf/nan path does not bloat the hot +// parser it is inlined into. template <typename UC> inline FASTFLOAT_CONSTEXPR14 bool fastfloat_strncasecmp(UC const *actual_mixedcase, UC const *expected_lowercase, size_t length) { - uint64_t mask{0}; - FASTFLOAT_IF_CONSTEXPR17(sizeof(UC) == 1) { mask = 0x2020202020202020; } - else FASTFLOAT_IF_CONSTEXPR17(sizeof(UC) == 2) { - mask = 0x0020002000200020; - } - else FASTFLOAT_IF_CONSTEXPR17(sizeof(UC) == 4) { - mask = 0x0000002000000020; - } - else { - return false; - } - - if (cpp20_and_in_constexpr()) { - for (size_t i = 0; i < length; i++) { - if ((actual_mixedcase[i] | 32) != expected_lowercase[i]) { - return false; - } + for (size_t i = 0; i < length; ++i) { + if ((actual_mixedcase[i] | 32) != expected_lowercase[i]) { + return false; } - return true; - } else { - uint64_t val1{0}, val2{0}; - size_t sz{8 / (sizeof(UC))}; - for (size_t i = 0; i < length; i += sz) { - val1 = val2 = 0; - sz = sz < (length - i) ? sz : length - i; - ::memcpy(&val1, actual_mixedcase + i, sz * sizeof(UC)); - ::memcpy(&val2, expected_lowercase + i, sz * sizeof(UC)); - val1 |= mask; - val2 |= mask; - if (val1 != val2) { - return false; - } - } - return true; } + return true; } #ifndef FLT_EVAL_METHOD
diff --git a/include/fast_float/parse_number.h b/include/fast_float/parse_number.h index 14d077b..febfdc7 100644 --- a/include/fast_float/parse_number.h +++ b/include/fast_float/parse_number.h
@@ -35,7 +35,7 @@ ++first; } if (last - first >= 3) { - if (fastfloat_strncasecmp3(first, str_const_nan<UC>())) { + if (fastfloat_strncasecmp(first, str_const_nan<UC>(), 3)) { answer.ptr = (first += 3); value = minusSign ? -std::numeric_limits<T>::quiet_NaN() : std::numeric_limits<T>::quiet_NaN(); @@ -54,9 +54,9 @@ } return answer; } - if (fastfloat_strncasecmp3(first, str_const_inf<UC>())) { + if (fastfloat_strncasecmp(first, str_const_inf<UC>(), 3)) { if ((last - first >= 8) && - fastfloat_strncasecmp5(first + 3, str_const_inf<UC>() + 3)) { + fastfloat_strncasecmp(first + 3, str_const_inf<UC>() + 3, 5)) { answer.ptr = first + 8; } else { answer.ptr = first + 3;