reuse pre-allocated thread locks in critical memoryview creation path

Cython/Utility/MemoryView.pyx

@@ -300,6 +300,24 @@ cdef void *align_pointer(void *memory, size_t alignment) nogil:
 
     return <void *> aligned_p
 
+
+# pre-allocate thread locks for reuse
+## note that this could be implemented in a more beautiful way in "normal" Cython,
+## but this code gets merged into the user module and not everything works there.
+DEF THREAD_LOCKS_PREALLOCATED = 8
+cdef int _thread_locks_used = 0
+cdef PyThread_type_lock[THREAD_LOCKS_PREALLOCATED] _thread_locks = [
+    PyThread_allocate_lock(),
+    PyThread_allocate_lock(),
+    PyThread_allocate_lock(),
+    PyThread_allocate_lock(),
+    PyThread_allocate_lock(),
+    PyThread_allocate_lock(),
+    PyThread_allocate_lock(),
+    PyThread_allocate_lock(),
+]
+
+
 @cname('__pyx_memoryview')
 cdef class memoryview(object):
 
@@ -325,8 +343,13 @@ cdef class memoryview(object):
                 (<__pyx_buffer *> &self.view).obj = Py_None
                 Py_INCREF(Py_None)
 
+        global _thread_locks_used
+        if _thread_locks_used < THREAD_LOCKS_PREALLOCATED:
+            self.lock = _thread_locks[_thread_locks_used]
+            _thread_locks_used += 1
+        if self.lock is NULL:
             self.lock = PyThread_allocate_lock()
-        if self.lock == NULL:
+            if self.lock is NULL:
                 raise MemoryError
 
         if flags & PyBUF_FORMAT:
@@ -342,7 +365,17 @@ cdef class memoryview(object):
         if self.obj is not None:
             __Pyx_ReleaseBuffer(&self.view)
 
+        cdef int i
+        global _thread_locks_used
         if self.lock != NULL:
+            for i in range(_thread_locks_used):
+                if _thread_locks[i] is self.lock:
+                    _thread_locks_used -= 1
+                    if i != _thread_locks_used:
+                        _thread_locks[i], _thread_locks[_thread_locks_used] = (
+                            _thread_locks[_thread_locks_used], _thread_locks[i])
+                    break
+            else:
                 PyThread_free_lock(self.lock)
 
     cdef char *get_item_pointer(memoryview self, object index) except NULL: