speed up (x % pyint)

CHANGES.rst

@@ -17,7 +17,8 @@ Features added
 
 * Adding/subtracting constant Python floats and small integers is faster.
 
-* Binary and/or/xor operations with small constant Python integers are faster.
+* Binary and/or/xor operations and modulus with small constant Python integers
+  are faster.
 
 Bugs fixed
 ----------

Cython/Compiler/Optimize.py

@@ -2811,6 +2811,13 @@ class OptimizeBuiltinCalls(Visitor.NodeRefCleanupMixin,
     def _handle_simple_method_object___xor__(self, node, function, args, is_unbound_method):
         return self._optimise_num_binop('Xor', node, function, args, is_unbound_method)
 
+    def _handle_simple_method_object___mod__(self, node, function, args, is_unbound_method):
+        if len(args) != 2 or not isinstance(args[1], ExprNodes.IntNode):
+            return node
+        if not args[1].has_constant_result() or not (2 <= args[1].constant_result <= 2**30):
+            return node
+        return self._optimise_num_binop('Remainder', node, function, args, is_unbound_method)
+
     def _handle_simple_method_float___add__(self, node, function, args, is_unbound_method):
         return self._optimise_num_binop('Add', node, function, args, is_unbound_method)
 

Cython/Utility/Optimize.c

@@ -493,28 +493,32 @@ static PyObject* __Pyx_PyInt_{{op}}{{order}}(PyObject *op1, PyObject *op2, long
 
 #if CYTHON_COMPILING_IN_CPYTHON
 {{py: pyval, ival = ('op2', 'b') if order == 'CObj' else ('op1', 'a') }}
-{{py: c_op = {'Add': '+', 'Subtract': '-', 'Or': '|', 'Xor': '^', 'And': '&'}[op] }}
-
-static PyObject* __Pyx_PyInt_{{op}}{{order}}(PyObject *op1, PyObject *op2, long intval, int inplace) {
-    const long {{'a' if order == 'CObj' else 'b'}} = intval;
+{{py: c_op = {'Add': '+', 'Subtract': '-', 'Remainder': '%', 'Or': '|', 'Xor': '^', 'And': '&'}[op] }}
 
+static PyObject* __Pyx_PyInt_{{op}}{{order}}(PyObject *op1, PyObject *op2, CYTHON_UNUSED long intval, int inplace) {
     #if PY_MAJOR_VERSION < 3
     if (likely(PyInt_CheckExact({{pyval}}))) {
+        const long {{'a' if order == 'CObj' else 'b'}} = intval;
         {{if c_op in '+-'}}
         long x;
         {{endif}}
         long {{ival}} = PyInt_AS_LONG({{pyval}});
 
-        {{if c_op not in '+-'}}
-        // binary operators are safe, no overflow
-        return PyInt_FromLong(a {{c_op}} b);
-
-        {{else}}
+        {{if c_op in '+-'}}
             // adapted from intobject.c in Py2.7:
             // casts in the line below avoid undefined behaviour on overflow
             x = (long)((unsigned long)a {{c_op}} b);
             if (likely((x^a) >= 0 || (x^{{ '~' if op == 'Subtract' else '' }}b) >= 0))
                 return PyInt_FromLong(x);
+
+        {{else}}
+            {{if c_op == '%'}}
+            // modulus with differing signs isn't safely portable
+            if (unlikely({{ival}} < 0))
+                return (inplace ? PyNumber_InPlace{{op}} : PyNumber_{{op}})(op1, op2);
+            {{endif}}
+            // other operations are safe, no overflow
+            return PyInt_FromLong(a {{c_op}} b);
         {{endif}}
 
         return PyLong_Type.tp_as_number->nb_{{op.lower()}}(op1, op2);
@@ -523,11 +527,20 @@ static PyObject* __Pyx_PyInt_{{op}}{{order}}(PyObject *op1, PyObject *op2, long
 
     #if PY_MAJOR_VERSION >= 3 && CYTHON_USE_PYLONG_INTERNALS
     if (likely(PyLong_CheckExact({{pyval}}))) {
+        const long {{'a' if order == 'CObj' else 'b'}} = intval;
         long {{ival}};
         switch (Py_SIZE({{pyval}})) {
+            {{if c_op != '%'}}
             case -1: {{ival}} = -(sdigit)((PyLongObject*){{pyval}})->ob_digit[0]; break;
+            {{endif}}
             case  0: {{ival}} = 0; break;
             case  1: {{ival}} = ((PyLongObject*){{pyval}})->ob_digit[0]; break;
+            case  2:
+                if (8 * sizeof(long) > 2 * PyLong_SHIFT) {
+                    {{ival}} = (long) ((((unsigned long)((PyLongObject*){{pyval}})->ob_digit[1]) << PyLong_SHIFT) | ((PyLongObject*){{pyval}})->ob_digit[0]);
+                    break;
+                }
+                // fall through if two platform digits don't fit into a long
             default: return PyLong_Type.tp_as_number->nb_{{op.lower()}}(op1, op2);
         }
         return PyLong_FromLong(a {{c_op}} b);
@@ -536,6 +549,7 @@ static PyObject* __Pyx_PyInt_{{op}}{{order}}(PyObject *op1, PyObject *op2, long
 
     {{if c_op in '+-'}}
     if (PyFloat_CheckExact({{pyval}})) {
+        const long {{'a' if order == 'CObj' else 'b'}} = intval;
         double result;
         double {{ival}} = PyFloat_AS_DOUBLE({{pyval}});
         // copied from floatobject.c in Py3.5:

tests/run/addop.pyx

 cimport cython
 
 
+def bigint(x):
+    print(str(x).rstrip('L'))
+
+
 def mixed_test():
     """
     >>> mixed_test()
@@ -29,6 +33,10 @@ def add_x_1(x):
     2
     >>> add_x_1(-1)
     0
+    >>> bigint(2**50 + 1)
+    1125899906842625
+    >>> bigint(add_x_1(2**50))
+    1125899906842625
     >>> add_x_1(1.5)
     2.5
     >>> add_x_1(-1.5)
@@ -73,13 +81,17 @@ def add_x_large(x):
     -1073741824.0
     >>> add_x_large(2**30 + 1)
     2147483649
+    >>> bigint(2**50 + 1 + 2**30)
+    1125900980584449
+    >>> bigint(add_x_large(2**50 + 1))
+    1125900980584449
     >>> 2**31 + 2**30
     3221225472
     >>> add_x_large(2**31)
     3221225472
-    >>> print(2**66 + 2**30)
+    >>> bigint(2**66 + 2**30)
     73786976295911948288
-    >>> print(add_x_large(2**66))
+    >>> bigint(add_x_large(2**66))
     73786976295911948288
     >>> try: add_x_large("abc")
     ... except TypeError: pass
@@ -96,6 +108,10 @@ def add_1_x(x):
     2
     >>> add_1_x(-1)
     0
+    >>> bigint(2**50 + 1)
+    1125899906842625
+    >>> bigint(add_1_x(2**50))
+    1125899906842625
     >>> add_1_x(1.5)
     2.5
     >>> add_1_x(-1.5)

tests/run/modop.pyx

 import sys
-if sys.version_info[0] < 3:
-    __doc__ = u"""
-    >>> modptr()
-    'spameggs'
-    """
+
 
 def modobj(obj2, obj3):
     """
@@ -15,6 +11,73 @@ def modobj(obj2, obj3):
     obj1 = obj2 % obj3
     return obj1
 
+
+def mod_obj_10(int2):
+    """
+    >>> 0 % 10
+    0
+    >>> mod_obj_10(0)
+    0
+    >>> 1 % 10
+    1
+    >>> mod_obj_10(1)
+    1
+    >>> 9 % 10
+    9
+    >>> mod_obj_10(9)
+    9
+    >>> 10 % 10
+    0
+    >>> mod_obj_10(10)
+    0
+    >>> 10002 % 10
+    2
+    >>> mod_obj_10(10002)
+    2
+    >>> int((2**50) % 10)
+    4
+    >>> int(mod_obj_10(2**50))
+    4
+    >>> int((2**200) % 10)
+    6
+    >>> int(mod_obj_10(2**200))
+    6
+    >>> (-1) % 10
+    9
+    >>> mod_obj_10(-1)
+    9
+    >>> (-10) % 10
+    0
+    >>> mod_obj_10(-10)
+    0
+    >>> (-12) % 10
+    8
+    >>> mod_obj_10(-12)
+    8
+    """
+    int1 = int2 % 10
+    return int1
+
+
+def mod_obj_m2(int2):
+    """
+    >>> 0 % -2
+    0
+    >>> mod_obj_m2(0)
+    0
+    >>> 1 % -2
+    -1
+    >>> mod_obj_m2(1)
+    -1
+    >>> 9 % -2
+    -1
+    >>> mod_obj_m2(9)
+    -1
+    """
+    int1 = int2 % -2
+    return int1
+
+
 def modint(int int2, int int3):
     """
     >>> modint(9,2)
@@ -24,7 +87,12 @@ def modint(int int2, int int3):
     int1 = int2 % int3
     return int1
 
+
 def modptr():
+    """
+    >>> print(modptr() if sys.version_info[0] < 3 else 'spameggs')
+    spameggs
+    """
     cdef char *str2, *str3
     str2 = "spam%s"
     str3 = "eggs"

tests/run/subop.pyx

 cimport cython
 
 
+def bigint(x):
+    print(str(x).rstrip('L'))
+
+
 def mixed_test():
     """
     >>> mixed_test()
@@ -44,6 +48,10 @@ def sub_x_1(x):
     0
     >>> sub_x_1(-1)
     -2
+    >>> bigint(2**50 - 1)
+    1125899906842623
+    >>> bigint(sub_x_1(2**50))
+    1125899906842623
     >>> sub_x_1(1.5)
     0.5
     >>> sub_x_1(-1.5)
@@ -63,6 +71,10 @@ def sub_x_1f(x):
     0.0
     >>> sub_x_1f(-1)
     -2.0
+    >>> 2**50 - 1.0
+    1125899906842623.0
+    >>> sub_x_1f(2**50)
+    1125899906842623.0
     >>> sub_x_1f(1.5)
     0.5
     >>> sub_x_1f(-1.5)
@@ -82,6 +94,10 @@ def sub_x_large(x):
     -1073741823
     >>> sub_x_large(-1)
     -1073741825
+    >>> bigint(2**50 - 2**30)
+    1125898833100800
+    >>> bigint(sub_x_large(2**50))
+    1125898833100800
     >>> sub_x_large(2.0**30)
     0.0
     >>> sub_x_large(2.0**30 + 1)
@@ -107,6 +123,10 @@ def sub_1_x(x):
     2
     >>> sub_1_x(1)
     0
+    >>> bigint(1 - 2**50)
+    -1125899906842623
+    >>> bigint(sub_1_x(2**50))
+    -1125899906842623
     >>> sub_1_x(1.5)
     -0.5
     >>> sub_1_x(-1.5)