gh-155742: Optimize json float parsing

Include/internal/pycore_floatobject.h

@@ -41,6 +41,9 @@ extern double _Py_parse_inf_or_nan(const char *p, char **endptr);
 
 extern int _Py_convert_int_to_double(PyObject **v, double *dbl);
 
+// Export for '_json' shared extension.
+PyAPI_FUNC(PyObject*) _PyFloat_FromString(const char *str, Py_ssize_t len);
+
 /* Should match endianness of the platform in most (all?) cases. */
 
 #ifdef DOUBLE_IS_BIG_ENDIAN_IEEE754

Modules/_json.c

@@ -12,6 +12,7 @@
 #include "pycore_ceval.h"         // _Py_EnterRecursiveCall()
 #include "pycore_critical_section.h" // Py_BEGIN_CRITICAL_SECTION_SEQUENCE_FAST()
 #include "pycore_dict.h"          // _PyDict_SetItem_Take2()
+#include "pycore_floatobject.h"   // _PyFloat_FromString()
 #include "pycore_list.h"          // _PyList_AppendTakeRef()
 #include "pycore_global_strings.h" // _Py_ID()
 #include "pycore_pyerrors.h"      // _PyErr_FormatNote
@@ -997,7 +998,6 @@ _match_number_unicode(PyScannerObject *s, PyObject *pystr, Py_ssize_t start, Py_
     Py_ssize_t idx = start;
     int is_float = 0;
     PyObject *rval;
-    PyObject *numstr = NULL;
     PyObject *custom_func;
 
     str = PyUnicode_DATA(pystr);
@@ -1064,32 +1064,39 @@ _match_number_unicode(PyScannerObject *s, PyObject *pystr, Py_ssize_t start, Py_
 
     if (custom_func) {
         /* copy the section we determined to be a number */
-        numstr = PyUnicode_FromKindAndData(kind,
-                                           (char*)str + kind * start,
-                                           idx - start);
-        if (numstr == NULL)
+        PyObject *numstr = PyUnicode_FromKindAndData(kind,
+                                                     (char*)str + kind * start,
+                                                     idx - start);
+        if (numstr == NULL) {
             return NULL;
+        }
         rval = PyObject_CallOneArg(custom_func, numstr);
+        Py_DECREF(numstr);
     }
     else {
-        Py_ssize_t i, n;
-        char *buf;
         /* Straight conversion to ASCII, to avoid costly conversion of
            decimal unicode digits (which cannot appear here) */
-        n = idx - start;
-        numstr = PyBytes_FromStringAndSize(NULL, n);
-        if (numstr == NULL)
+        Py_ssize_t n = idx - start;
+        char *buf = PyMem_Malloc(n + 1);
+        if (buf == NULL) {
+            PyErr_NoMemory();
             return NULL;
-        buf = PyBytes_AS_STRING(numstr);
-        for (i = 0; i < n; i++) {
-            buf[i] = (char) PyUnicode_READ(kind, str, i + start);
         }
-        if (is_float)
-            rval = PyFloat_FromString(numstr);
-        else
+
+        for (Py_ssize_t i = 0; i < n; i++) {
+            Py_UCS4 ch = PyUnicode_READ(kind, str, i + start);
+            assert(ch <= 127);
+            buf[i] = (char)ch;
+        }
+        buf[n] = '\0';
+        if (is_float) {
+            rval = _PyFloat_FromString(buf, n);
+        }
+        else {
             rval = PyLong_FromString(buf, NULL, 10);
+        }
+        PyMem_Free(buf);
     }
-    Py_DECREF(numstr);
     *next_idx_ptr = idx;
     return rval;
 }

Objects/floatobject.c

@@ -137,20 +137,18 @@ PyFloat_FromDouble(double fval)
 }
 
 static PyObject *
-float_from_string_inner(const char *s, Py_ssize_t len, void *obj)
+float_from_string_inner(const char *s, Py_ssize_t len, void *data)
 {
     double x;
+    const char *orig_s = s;
     const char *end;
     const char *last = s + len;
     /* strip leading whitespace */
     while (s < last && Py_ISSPACE(*s)) {
         s++;
     }
     if (s == last) {
-        PyErr_Format(PyExc_ValueError,
-                     "could not convert string to float: "
-                     "%R", obj);
-        return NULL;
+        goto error;
     }
 
     /* strip trailing whitespace */
@@ -163,19 +161,49 @@ float_from_string_inner(const char *s, Py_ssize_t len, void *obj)
      * fine. */
     x = PyOS_string_to_double(s, (char **)&end, NULL);
     if (end != last) {
-        PyErr_Format(PyExc_ValueError,
-                     "could not convert string to float: "
-                     "%R", obj);
-        return NULL;
+        goto error;
     }
     else if (x == -1.0 && PyErr_Occurred()) {
         return NULL;
     }
     else {
         return PyFloat_FromDouble(x);
     }
+
+error:
+    // Label followed by a declaration is a C23 extension, so use a sub-scope
+    {
+        PyObject *obj = (PyObject*)data;
+        if (obj == NULL) {
+            obj = PyBytes_FromStringAndSize(orig_s, len);
+            if (obj == NULL) {
+                return NULL;
+            }
+        }
+        else {
+            Py_INCREF(obj);
+        }
+        PyErr_Format(PyExc_ValueError,
+                     "could not convert string to float: "
+                     "%R", obj);
+        Py_DECREF(obj);
+    }
+    return NULL;
+}
+
+
+// Internal API similar to PyFloat_FromString() but doesn't require a Python
+// object. str[len] must be the NUL byte.
+PyObject*
+_PyFloat_FromString(const char *str, Py_ssize_t len)
+{
+    assert(len >= 1);
+    assert(str[len] == '\0');
+    return _Py_string_to_number_with_underscores(str, len, "float", NULL, NULL,
+                                                 float_from_string_inner);
 }
 
+
 PyObject *
 PyFloat_FromString(PyObject *v)
 {

Python/pystrtod.c

@@ -395,9 +395,19 @@ _Py_string_to_number_with_underscores(
 
   error:
     PyMem_Free(dup);
+    if (obj == NULL) {
+        obj = PyBytes_FromStringAndSize(s, orig_len);
+        if (obj == NULL) {
+            return NULL;
+        }
+    }
+    else {
+        Py_INCREF(obj);
+    }
     PyErr_Format(PyExc_ValueError,
                  "could not convert string to %s: "
                  "%R", what, obj);
+    Py_DECREF(obj);
     return NULL;
 }