tweak 2**N optimisation a bit more in Py3

Cython/Utility/Optimize.c

@@ -435,6 +435,7 @@ bad:
 static PyObject* __Pyx__PyNumber_PowerOf2(PyObject *two, PyObject *exp, PyObject *none, int inplace); /*proto*/
 
 /////////////// PyNumberPow2 ///////////////
+//@requires: TypeConversion.c::PyLongInternals
 
 static PyObject* __Pyx__PyNumber_PowerOf2(PyObject *two, PyObject *exp, PyObject *none, int inplace) {
 // in CPython, 1<<N is substantially faster than 2**N
@@ -442,7 +443,17 @@ static PyObject* __Pyx__PyNumber_PowerOf2(PyObject *two, PyObject *exp, PyObject
 #if CYTHON_COMPILING_IN_CPYTHON
     Py_ssize_t shiftby;
     if (likely(PyLong_CheckExact(exp))) {
+        #if PY_MAJOR_VERSION >= 3 && CYTHON_USE_PYLONG_INTERNALS
+        switch (Py_SIZE(exp)) {
+            case  0: shiftby = 0; break;
+            case  1: shiftby = ((PyLongObject*)exp)->ob_digit[0]; break;
+            default:
+                if (unlikely(Py_SIZE(exp) < 0)) goto fallback;
+                shiftby = PyLong_AsSsize_t(exp); break;
+        }
+        #else
         shiftby = PyLong_AsSsize_t(exp);
+        #endif
 #if PY_MAJOR_VERSION < 3
     } else if (likely(PyInt_CheckExact(exp))) {
         shiftby = PyInt_AsLong(exp);

Cython/Utility/TypeConversion.c

@@ -150,6 +150,7 @@ bad:
 #endif
 
 /////////////// TypeConversions ///////////////
+//@requires: PyLongInternals
 
 /* Type Conversion Functions */
 
@@ -279,11 +280,6 @@ static CYTHON_INLINE PyObject* __Pyx_PyNumber_Int(PyObject* x) {
   return res;
 }
 
-#if CYTHON_COMPILING_IN_CPYTHON && PY_MAJOR_VERSION >= 3
- #if CYTHON_USE_PYLONG_INTERNALS
-  #include "longintrepr.h"
- #endif
-#endif
 static CYTHON_INLINE Py_ssize_t __Pyx_PyIndex_AsSsize_t(PyObject* b) {
   Py_ssize_t ival;
   PyObject *x;
@@ -504,18 +500,23 @@ static CYTHON_INLINE PyObject* {{TO_PY_FUNCTION}}({{TYPE}} value) {
     }
 
 
+/////////////// PyLongInternals ///////////////
+
+#if CYTHON_COMPILING_IN_CPYTHON && PY_MAJOR_VERSION >= 3
+ #if CYTHON_USE_PYLONG_INTERNALS
+  #include "longintrepr.h"
+ #endif
+#endif
+
+
 /////////////// CIntFromPy.proto ///////////////
 
 static CYTHON_INLINE {{TYPE}} {{FROM_PY_FUNCTION}}(PyObject *);
 
 /////////////// CIntFromPy ///////////////
 //@requires: CIntFromPyVerify
+//@requires: PyLongInternals
 
-#if CYTHON_COMPILING_IN_CPYTHON && PY_MAJOR_VERSION >= 3
- #if CYTHON_USE_PYLONG_INTERNALS
-  #include "longintrepr.h"
- #endif
-#endif
 static CYTHON_INLINE {{TYPE}} {{FROM_PY_FUNCTION}}(PyObject *x) {
     const {{TYPE}} neg_one = ({{TYPE}}) -1, const_zero = 0;
     const int is_unsigned = neg_one > const_zero;