coffee/fast_float
1
0
Fork 0
mirror of https://github.com/fastfloat/fast_float.git synced 2026-07-30 08:16:25 +08:00
Code Issues Packages Projects Releases Wiki Activity

Merge pull request #398 from redis-performance/exp062-063-combo

Browse Source 
Enable the 4-digit SWAR follow-up on GCC + skip the rounding-mode probe for long mantissas
This commit is contained in:
Daniel Lemire 2026-07-13 15:13:16 -04:00 committed by GitHub
commit f6df0f2917
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 20 additions and 17 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

8
include/fast_float/ascii_number.h
View File

@ -268,10 +268,9 @@ loop_parse_if_eight_digits(char const *&p, char const *const pend,
}
// Consume a remaining 4-7 digit run in a single SWAR step instead of
// byte-by-byte (reuses the existing 4-digit helpers). The parsed result is
// identical either way. Gated to clang: on gcc the extra 4-digit check
// regresses inputs whose remainder is shorter than 4 digits (it becomes pure
// overhead there); clang does not show that.
#if defined(__clang__)
// identical either way. Historically gated to clang because gcc regressed on
// short remainders, but that verdict predates the span-elision restructure;
// with the leaner hot path the 4-digit step now wins on gcc as well.
if ((pend - p) >= 4) {
uint32_t const val4 = read4_to_u32(p);
if (is_made_of_four_digits_fast(val4)) {
@ -280,7 +279,6 @@ loop_parse_if_eight_digits(char const *&p, char const *const pend,
p += 4;
}
}
#endif
}
enum class parse_error {

11
include/fast_float/parse_number.h
View File

@ -198,7 +198,14 @@ clinger_fast_path_impl(uint64_t mantissa, int64_t exponent, bool is_negative,
// We proceed optimistically, assuming that detail::rounds_to_nearest()
// returns true.
if (binary_format<T>::min_exponent_fast_path() <= exponent &&
exponent <= binary_format<T>::max_exponent_fast_path()) {
exponent <= binary_format<T>::max_exponent_fast_path() &&
mantissa <= binary_format<T>::max_mantissa_fast_path()) {
// The mantissa bound above is a necessary condition for BOTH branches
// below: the rounding-mode-dependent branch checks the tighter
// max_mantissa_fast_path(exponent) <= max_mantissa_fast_path(). Testing
// it before detail::rounds_to_nearest() spares long-mantissa inputs
// (which can never take the fast path) the volatile-float probe.
//
// Unfortunately, the conventional Clinger's fast path is only possible
// when the system rounds to the nearest float.
//
@ -209,7 +216,6 @@ clinger_fast_path_impl(uint64_t mantissa, int64_t exponent, bool is_negative,
if (!cpp20_and_in_constexpr() && detail::rounds_to_nearest()) {
// We have that fegetround() == FE_TONEAREST.
// Next is Clinger's fast path.
if (mantissa <= binary_format<T>::max_mantissa_fast_path()) {
value = T(mantissa);
if (exponent < 0) {
value = value / binary_format<T>::exact_power_of_ten(-exponent);
@ -220,7 +226,6 @@ clinger_fast_path_impl(uint64_t mantissa, int64_t exponent, bool is_negative,
value = -value;
}
return true;
}
} else {
// We do not have that fegetround() == FE_TONEAREST.
// Next is a modified Clinger's fast path, inspired by Jakub Jelínek's