From 9b3d9f7b1a4ddb59449166800c5281ab28cd4fbc Mon Sep 17 00:00:00 2001
From: HedgehogInTheCPP <a.rainman@gmail.com>
Date: Sat, 29 Aug 2026 17:33:00 +0300
Subject: [PATCH] add fastfloat_noinline from
 https://github.com/fastfloat/fast_float/pull/403 my fork don't need any
 additional from this PR because it's already simd optimized more strictly.

---
 include/fast_float/digit_comparison.h | 9 +++++----
 include/fast_float/float_common.h     | 8 +++++++-
 include/fast_float/parse_number.h     | 2 +-
 3 files changed, 13 insertions(+), 6 deletions(-)

diff --git a/include/fast_float/digit_comparison.h b/include/fast_float/digit_comparison.h
index 710aa6e..435a2c8 100644
--- a/include/fast_float/digit_comparison.h
+++ b/include/fast_float/digit_comparison.h
@@ -379,8 +379,9 @@ parse_mantissa(bigint &result, const parsed_number_string_t<UC> &num) noexcept {
 }
 
 template <typename T>
-inline FASTFLOAT_CONSTEXPR20 adjusted_mantissa positive_digit_comp(
-    bigint &bigmant, adjusted_mantissa am, am_pow_t const exponent) noexcept {
+fastfloat_really_inline FASTFLOAT_CONSTEXPR20 adjusted_mantissa
+positive_digit_comp(bigint &bigmant, adjusted_mantissa am,
+                    am_pow_t const exponent) noexcept {
   FASTFLOAT_ASSERT(bigmant.pow10(exponent));
   bool truncated;
   am.mantissa = bigmant.hi64(truncated);
@@ -407,7 +408,7 @@ inline FASTFLOAT_CONSTEXPR20 adjusted_mantissa positive_digit_comp(
 // we then need to scale by `2^(f- e)`, and then the two significant digits
 // are of the same magnitude.
 template <typename T>
-inline FASTFLOAT_CONSTEXPR20 adjusted_mantissa
+fastfloat_really_inline FASTFLOAT_CONSTEXPR20 adjusted_mantissa
 negative_digit_comp(bigint &real_digits, adjusted_mantissa am,
                     am_pow_t const real_exp) noexcept {
   // get the value of `b`, rounded down, and get a bigint representation of b+h
@@ -472,7 +473,7 @@ negative_digit_comp(bigint &real_digits, adjusted_mantissa am,
 // the actual digits. we then compare the big integer representations
 // of both, and use that to direct rounding.
 template <typename T, typename UC>
-inline FASTFLOAT_CONSTEXPR20 adjusted_mantissa digit_comp(
+fastfloat_noinline FASTFLOAT_CONSTEXPR20 adjusted_mantissa digit_comp(
     parsed_number_string_t<UC> const &num, adjusted_mantissa am) noexcept {
   // remove the invalid exponent bias
   am.power2 -= invalid_am_bias;
diff --git a/include/fast_float/float_common.h b/include/fast_float/float_common.h
index 10cc861..6f74c25 100644
--- a/include/fast_float/float_common.h
+++ b/include/fast_float/float_common.h
@@ -230,10 +230,16 @@ static_assert(FASTFLOAT_X86_SIMD == 20 || FASTFLOAT_X86_SIMD == 42 ||
 #endif
 #endif
 
+
 #ifdef FASTFLOAT_VISUAL_STUDIO
 #define fastfloat_really_inline __forceinline
-#else
+#define fastfloat_noinline __declspec(noinline)
+#elif defined(__GNUC__) || defined(__clang__)
 #define fastfloat_really_inline inline __attribute__((always_inline))
+#define fastfloat_noinline __attribute__((noinline))
+#else
+#define fastfloat_really_inline inline
+#define fastfloat_noinline
 #endif
 
 // Branch-probability hint marking the rare slow-path branches as cold, so the
diff --git a/include/fast_float/parse_number.h b/include/fast_float/parse_number.h
index 548db8d..7faa3a2 100644
--- a/include/fast_float/parse_number.h
+++ b/include/fast_float/parse_number.h
@@ -331,7 +331,7 @@ from_chars_advanced(parsed_number_string_t<UC> const &pns, T &value) noexcept {
 // and the am.power2<0 digit_comp recompute, so both slow branches collapse to
 // one helper call.
 template <typename T, typename UC>
-FASTFLOAT_CONSTEXPR20 from_chars_result_t<UC>
+static fastfloat_noinline FASTFLOAT_CONSTEXPR20 from_chars_result_t<UC>
 parse_number_slow_path(UC const *first, UC const *last, T &value,
                        parse_options_t<UC> const options
 #ifndef FASTFLOAT_ONLY_POSITIVE_C_NUMBER_WO_INF_NAN