Enable the 4-digit SWAR follow-up on GCC + skip the rounding-mode probe for long mantissas

include/fast_float/ascii_number.h

@@ -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 {

include/fast_float/parse_number.h

@@ -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.
     //
@@ -209,18 +216,16 @@ 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);
-        } else {
-          value = value * binary_format<T>::exact_power_of_ten(exponent);
-        }
-        if (is_negative) {
-          value = -value;
-        }
-        return true;
+      value = T(mantissa);
+      if (exponent < 0) {
+        value = value / binary_format<T>::exact_power_of_ten(-exponent);
+      } else {
+        value = value * binary_format<T>::exact_power_of_ten(exponent);
+      }
+      if (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