gh-153740: Optimize PyFloat_Pack/Unpack2 using native _Float16

Lib/test/test_capi/test_float.py

@@ -226,6 +226,8 @@ def test_pack_unpack_roundtrip_for_nans(self):
                         value = unpack(data1, endian)
                         data2 = pack(size, value, endian)
                         self.assertTrue(math.isnan(value))
+                        self.assertEqual(math.copysign(1.0, value),
+                                         -1.0 if sign else 1.0)
                         self.assertEqual(data1, data2)
 
     @unittest.skipUnless(HAVE_IEEE_754, "requires IEEE 754")

Misc/NEWS.d/next/C_API/2026-07-28-04-29-52.gh-issue-153740.7_LMSA.rst

@@ -0,0 +1,3 @@
+If available on the platform, use native :c:type:`_Float16` in
+:c:func:`PyFloat_Pack2` and :c:func:`PyFloat_Unpack2` functions.  Patch by
+Sergey B Kirpichev.

Objects/floatobject.c

@@ -1895,6 +1895,31 @@ int
 PyFloat_Pack2(double x, char *data, int le)
 {
     unsigned char *p = (unsigned char *)data;
+#if HAVE_FLOAT16
+    /* Conversion can change NaNs type or alter payload.  Here we
+       just fallback to the generic code, instead of providing
+       workarounds as for single/double precision. */
+    if (!isnan(x)) {
+        _Float16 y = (_Float16)x;
+
+        if (isinf(y) && !isinf(x)) {
+            goto Overflow;
+        }
+
+        unsigned char s[sizeof(_Float16)];
+
+        memcpy(s, &y, sizeof(_Float16));
+        if ((_PY_FLOAT_LITTLE_ENDIAN && !le) || (_PY_FLOAT_BIG_ENDIAN && le)) {
+            p[1] = s[0];
+            p[0] = s[1];
+        }
+        else {
+            p[0] = s[0];
+            p[1] = s[1];
+        }
+        return 0;
+    }
+#endif
     unsigned char sign;
     int e;
     double f;
@@ -2090,6 +2115,24 @@ double
 PyFloat_Unpack2(const char *data, int le)
 {
     unsigned char *p = (unsigned char *)data;
+#if HAVE_FLOAT16
+    _Float16 x16;
+
+    if ((_PY_FLOAT_LITTLE_ENDIAN && !le) || (_PY_FLOAT_BIG_ENDIAN && le)) {
+        char buf[2];
+
+        buf[1] = p[0];
+        buf[0] = p[1];
+        memcpy(&x16, buf, 2);
+    }
+    else {
+        memcpy(&x16, p, 2);
+    }
+    if (!isnan(x16)) {
+        return x16;
+    }
+    /* Fallback to the generic code for NaNs, see PyFloat_Pack2(). */
+#endif
     unsigned char sign;
     int e;
     unsigned int f;

configure.ac

@@ -4469,6 +4469,25 @@ if test "$ac_cv_ffi_complex_double_supported" = "yes"; then
               [Defined if _Complex C type can be used with libffi.])
 fi
 
+# Check for native half-float type (_Float16).
+AC_CACHE_CHECK([for _Float16 support], [ac_cv_float16_supported],
+WITH_SAVE_ENV([
+CFLAGS="$CFLAGS -O0"
+AC_RUN_IFELSE([AC_LANG_SOURCE([[
+int main(void)
+{
+    _Float16 val = 1.0f16;
+    double d = 3.14;
+    val = d;
+    return 0;
+}
+]])], [ac_cv_float16_supported=yes],
+[ac_cv_float16_supported=no],
+[ac_cv_float16_supported=no])]))
+AS_VAR_IF([ac_cv_float16_supported], [yes],
+          [AC_DEFINE([HAVE_FLOAT16], [1],
+                     [Defined if _Float16 C type is supported])])
+
 dnl Check for libmpdec >= 2.5.0
 PKG_CHECK_MODULES([LIBMPDEC], [libmpdec >= 2.5.0], [have_mpdec=yes], [
   WITH_SAVE_ENV([

pyconfig.h.in

@@ -497,6 +497,9 @@
 /* Define if you have the 'ffi_prep_closure_loc' function. */
 #undef HAVE_FFI_PREP_CLOSURE_LOC
 
+/* Defined if _Float16 C type is supported */
+#undef HAVE_FLOAT16
+
 /* Define to 1 if you have the 'flock' function. */
 #undef HAVE_FLOCK