Skip to content

Speed up parsing: branch-free exponent, UTF-16 steps, clang parity #416

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
Algunenano wants to merge 8 commits into
base: main
Choose a base branch
from
Open

Speed up parsing: branch-free exponent, UTF-16 steps, clang parity #416

Changes from 1 commit
Commits
Show all changes
8 commits
Select commit Hold shift + click to select a range
3e852fe
Avoid branch mispredictions on the exponent sign and out-of-range res…
Algunenano Sep 23, 2026
18ee067
Apply clang-format
Algunenano Sep 23, 2026
0900eed
Inline the parser into from_chars under clang
Algunenano Sep 23, 2026
ef5797f
Skip the range check in compute_float after Clinger's range test unde…
Algunenano Sep 23, 2026
1eac92b
Parse the fraction tail and the exponent with one subtraction per digit
Algunenano Sep 23, 2026
a6777b6
Parse UTF-16 digits with a 4-unit step and combine digit pairs in SIMD
Algunenano Sep 24, 2026
8218446
Combine UTF-16 digit pairs on the characters and offset on the i side
Algunenano Sep 24, 2026
784cd67
Parse the integer part with one subtraction per digit
Algunenano Sep 24, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
Prev Previous commit
Next Next commit
Combine UTF-16 digit pairs on the characters and offset on the i side 
The '0' offset is linear, so it is subtracted once from i * 10^8, which is
ready early, instead of from each unit on the digit chain. Clang also
extracts both halves with one movq.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
  • Loading branch information
@Algunenano @claude
Algunenano and claude committed Sep 24, 2026
commit 8218446e2fd8d5c0c6a324b13f731032a350fc24
32 changes: 20 additions & 12 deletions include/fast_float/ascii_number.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -110,15 +110,18 @@ fastfloat_really_inline uint64_t simd_read8_to_u64(char16_t const *chars) {
FASTFLOAT_SIMD_RESTORE_WARNINGS
}

// Eight digit units to their value, combining digit pairs in the vector unit
// i followed by eight digit units, combining digit pairs in the vector unit
// (pmaddwd by 10,1 then 100,1): parse_eight_digits_unrolled on the packed
// bytes needs three 64-bit constants in general registers, and GCC 14/15
// spill a value on the digit chain to make room.
fastfloat_really_inline uint32_t simd_eight_digits_value(__m128i const data) {
fastfloat_really_inline uint64_t simd_append_eight_digits(uint64_t i,
__m128i const data) {
FASTFLOAT_SIMD_DISABLE_WARNINGS
__m128i const digits = _mm_sub_epi16(data, _mm_set1_epi16('0'));
// On the characters rather than the digits, with '0' * 11111111 taken off
// at the end: one step less on the digit chain (pairs stay below 628 and
// groups of four below 63328, in range for pmaddwd).
__m128i const pairs =
_mm_madd_epi16(digits, _mm_set_epi16(1, 10, 1, 10, 1, 10, 1, 10));
_mm_madd_epi16(data, _mm_set_epi16(1, 10, 1, 10, 1, 10, 1, 10));
__m128i const quads =
_mm_madd_epi16(_mm_packs_epi32(pairs, pairs),
_mm_set_epi16(1, 100, 1, 100, 1, 100, 1, 100));
Expand All @@ -128,7 +131,10 @@ fastfloat_really_inline uint32_t simd_eight_digits_value(__m128i const data) {
uint64_t both;
_mm_storel_epi64(reinterpret_cast<__m128i *>(&both), quads);
#endif
return uint32_t(both) * 10000 + uint32_t(both >> 32);
// The offset goes on the i side, which is ready early (clang otherwise
// subtracts it on the digit chain).
return (i * 100000000 - uint64_t('0') * 11111111) +
(uint32_t(both) * 10000 + uint32_t(both >> 32));
FASTFLOAT_SIMD_RESTORE_WARNINGS
}

Expand All @@ -149,17 +155,19 @@ fastfloat_really_inline uint64_t simd_read8_to_u64(char16_t const *chars) {
}

// See the SSE2 version.
fastfloat_really_inline uint32_t
simd_eight_digits_value(uint16x8_t const data) {
fastfloat_really_inline uint64_t
simd_append_eight_digits(uint64_t i, uint16x8_t const data) {
FASTFLOAT_SIMD_DISABLE_WARNINGS
static uint16_t const m10[8] = {10, 1, 10, 1, 10, 1, 10, 1};
static uint16_t const m100[4] = {100, 1, 100, 1};
uint16x8_t const digits = vsubq_u16(data, vdupq_n_u16('0'));
uint32x4_t const pairs = vpaddlq_u16(vmulq_u16(digits, vld1q_u16(m10)));
uint32x4_t const pairs = vpaddlq_u16(vmulq_u16(data, vld1q_u16(m10)));
uint32x2_t const quads =
vpaddl_u16(vmul_u16(vmovn_u32(pairs), vld1_u16(m100)));
uint64_t const both = vget_lane_u64(vreinterpret_u64_u32(quads), 0);
return uint32_t(both) * 10000 + uint32_t(both >> 32);
// The offset goes on the i side, which is ready early (clang otherwise
// subtracts it on the digit chain).
return (i * 100000000 - uint64_t('0') * 11111111) +
(uint32_t(both) * 10000 + uint32_t(both >> 32));
FASTFLOAT_SIMD_RESTORE_WARNINGS
}

Expand Down Expand Up @@ -239,7 +247,7 @@ simd_parse_if_eight_digits_unrolled(char16_t const *chars,
__m128i const t1 = _mm_cmpgt_epi16(t0, _mm_set1_epi16(-32759));

if (_mm_movemask_epi8(t1) == 0) {
i = i * 100000000 + simd_eight_digits_value(data);
i = simd_append_eight_digits(i, data);
return true;
} else
return false;
Expand All @@ -254,7 +262,7 @@ simd_parse_if_eight_digits_unrolled(char16_t const *chars,
uint16x8_t const mask = vcltq_u16(t0, vmovq_n_u16('9' - '0' + 1));

if (vminvq_u16(mask) == 0xFFFF) {
i = i * 100000000 + simd_eight_digits_value(data);
i = simd_append_eight_digits(i, data);
return true;
} else
return false;
Expand Down