Allow nested prange() & less warnings

Cython/Compiler/Nodes.py

@@ -6834,10 +6834,11 @@ class ParallelStatNode(StatNode, ParallelNode):
         self.analyse_sharing_attributes(env)
 
         if self.num_threads is not None:
-            if self.parent and self.parent.num_threads is not None:
+            if (self.parent and self.parent.num_threads is not None and not
+                                                    self.parent.is_prange):
                 error(self.pos,
                       "num_threads already declared in outer section")
-            elif self.parent:
+            elif self.parent and not self.parent.is_prange:
                 error(self.pos,
                       "num_threads must be declared in the parent parallel section")
             elif (self.num_threads.type.is_int and
@@ -7113,11 +7114,15 @@ class ParallelStatNode(StatNode, ParallelNode):
             begin_code = self.begin_of_parallel_block
             end_code = code
 
+            begin_code.putln("#ifdef _OPENMP")
             begin_code.put_ensure_gil(declare_gilstate=True)
             begin_code.putln("Py_BEGIN_ALLOW_THREADS")
+            begin_code.putln("#endif /* _OPENMP */")
 
+            end_code.putln("#ifdef _OPENMP")
             end_code.putln("Py_END_ALLOW_THREADS")
             end_code.put_release_ensured_gil()
+            end_code.putln("#endif /* _OPENMP */")
 
     def trap_parallel_exit(self, code, should_flush=False):
         """
@@ -7216,8 +7221,13 @@ class ParallelStatNode(StatNode, ParallelNode):
 
             temp_count += 1
 
+            invalid_value = entry.type.invalid_value()
+            if invalid_value:
+                init = ' = ' + invalid_value
+            else:
+                init = ''
             # Declare the parallel private in the outer block
-            c.putln("%s %s;" % (type_decl, temp_cname))
+            c.putln("%s %s%s;" % (type_decl, temp_cname, init))
 
             # Initialize before escaping
             code.putln("%s = %s;" % (temp_cname, private_cname))
@@ -7434,6 +7444,7 @@ class ParallelRangeNode(ParallelStatNode):
     start = stop = step = None
 
     is_prange = True
+    is_nested_prange = False
 
     nogil = None
     schedule = None
@@ -7544,6 +7555,16 @@ class ParallelRangeNode(ParallelStatNode):
         if self.nogil:
             env.nogil = was_nogil
 
+        self.is_nested_prange = self.parent and self.parent.is_prange
+        if self.is_nested_prange:
+            parent = self
+            while parent.parent and parent.parent.is_prange:
+                parent = parent.parent
+
+            parent.assignments.update(self.assignments)
+            parent.privates.update(self.privates)
+            parent.assigned_nodes.extend(self.assigned_nodes)
+
     def nogil_check(self, env):
         names = 'start', 'stop', 'step', 'target'
         nodes = self.start, self.stop, self.step, self.target
@@ -7670,7 +7691,10 @@ class ParallelRangeNode(ParallelStatNode):
         self.release_closure_privates(code)
 
     def generate_loop(self, code, fmt_dict):
-        code.putln("#ifdef _OPENMP")
+        if self.is_nested_prange:
+            code.putln("#if 0")
+        else:
+            code.putln("#ifdef _OPENMP")
 
         if not self.is_parallel:
             code.put("#pragma omp for")
@@ -7688,7 +7712,10 @@ class ParallelRangeNode(ParallelStatNode):
             # Initialize the GIL if needed for this thread
             self.begin_parallel_block(code)
 
-            code.putln("#ifdef _OPENMP")
+            if self.is_nested_prange:
+                code.putln("#if 0")
+            else:
+                code.putln("#ifdef _OPENMP")
             code.put("#pragma omp for")
 
         for entry, (op, lastprivate) in self.privates.iteritems():

Cython/Compiler/ParseTreeTransforms.py

@@ -1099,7 +1099,7 @@ class ParallelRangeTransform(CythonTransform, SkipDeclarations):
         if isinstance(newnode, Nodes.ParallelWithBlockNode):
             if self.state == 'parallel with':
                 error(node.manager.pos,
-                      "Closely nested parallel with blocks are disallowed")
+                      "Nested parallel with blocks are disallowed")
 
             self.state = 'parallel with'
             body = self.visit(node.body)

Cython/Compiler/TypeInference.py

@@ -196,10 +196,8 @@ class MarkAssignments(CythonTransform):
         self.parallel_block_stack.append(node)
 
         nested = nested or len(self.parallel_block_stack) > 2
-        if not self.parallel_errors and nested:
-
-            error(node.pos,
-                  "Parallel nesting not supported due to bugs in gcc 4.5")
+        if not self.parallel_errors and nested and not node.is_prange:
+            error(node.pos, "Only prange() may be nested")
             self.parallel_errors = True
 
         if node.is_prange:

tests/errors/e_cython_parallel.pyx

@@ -145,6 +145,10 @@ cdef int chunksize():
 for i in prange(10, nogil=True, schedule='static', chunksize=chunksize()):
     pass
 
+with nogil, cython.parallel.parallel():
+    with cython.parallel.parallel():
+        pass
+
 _ERRORS = u"""
 e_cython_parallel.pyx:3:8: cython.parallel.parallel is not a module
 e_cython_parallel.pyx:4:0: No such directive: cython.parallel.something
@@ -157,7 +161,7 @@ c_cython_parallel.pyx:21:29: The parallel section may only be used without the G
 e_cython_parallel.pyx:27:10: target may not be a Python object as we don't have the GIL
 e_cython_parallel.pyx:30:9: Can only iterate over an iteration variable
 e_cython_parallel.pyx:33:10: Must be of numeric type, not int *
-e_cython_parallel.pyx:36:33: Closely nested parallel with blocks are disallowed
+e_cython_parallel.pyx:36:33: Nested parallel with blocks are disallowed
 e_cython_parallel.pyx:39:12: The parallel directive must be called
 e_cython_parallel.pyx:45:10: local variable 'y' referenced before assignment
 e_cython_parallel.pyx:55:9: local variable 'y' referenced before assignment
@@ -166,8 +170,6 @@ e_cython_parallel.pyx:62:36: cython.parallel.parallel() does not take positional
 e_cython_parallel.pyx:65:36: Invalid keyword argument: invalid
 e_cython_parallel.pyx:73:12: Yield not allowed in parallel sections
 e_cython_parallel.pyx:77:16: Yield not allowed in parallel sections
-e_cython_parallel.pyx:82:19: Parallel nesting not supported due to bugs in gcc 4.5
-e_cython_parallel.pyx:87:23: Parallel nesting not supported due to bugs in gcc 4.5
 e_cython_parallel.pyx:97:19: Cannot assign to private of outer parallel block
 e_cython_parallel.pyx:98:19: Cannot assign to private of outer parallel block
 e_cython_parallel.pyx:104:6: Reductions not allowed for parallel blocks
@@ -180,4 +182,5 @@ e_cython_parallel.pyx:133:42: Must provide schedule with chunksize
 e_cython_parallel.pyx:136:62: Chunksize must not be negative
 e_cython_parallel.pyx:139:62: Chunksize not valid for the schedule runtime
 e_cython_parallel.pyx:145:70: Calling gil-requiring function not allowed without gil
+e_cython_parallel.pyx:149:33: Nested parallel with blocks are disallowed
 """

tests/run/sequential_parallel.pyx

@@ -46,7 +46,6 @@ def test_descending_prange():
 
     return sum
 
-'''
 def test_propagation():
     """
     >>> test_propagation()
@@ -61,12 +60,10 @@ def test_propagation():
 
     with nogil, cython.parallel.parallel():
         for x in prange(10):
-            with cython.parallel.parallel():
-                for y in prange(10):
-                    sum2 += y
+            for y in prange(10):
+                sum2 += y
 
     return i, j, x, y, sum1, sum2
-'''
 
 def test_unsigned_operands():
     """
@@ -349,7 +346,6 @@ def test_prange_continue():
 
     free(p)
 
-'''
 def test_nested_break_continue():
     """
     >>> test_nested_break_continue()
@@ -371,16 +367,14 @@ def test_nested_break_continue():
 
     print i, j, result1, result2
 
-    with nogil, cython.parallel.parallel():
-        for i in prange(10, num_threads=2, schedule='static'):
-            with cython.parallel.parallel():
-                if i == 8:
-                    break
-                else:
-                    continue
+    with nogil, cython.parallel.parallel(num_threads=2):
+        for i in prange(10, schedule='static'):
+            if i == 8:
+                break
+            else:
+                continue
 
     print i
-'''
 
 cdef int parallel_return() nogil:
     cdef int i
@@ -400,11 +394,10 @@ def test_return():
     """
     print parallel_return()
 
-'''
 def test_parallel_exceptions():
     """
     >>> test_parallel_exceptions()
-    ('I am executed first', 0)
+    I am executed first
     ('propagate me',) 0
     """
     cdef int i, j, sum = 0
@@ -422,11 +415,10 @@ def test_parallel_exceptions():
                 sum += i
             finally:
                 with gil:
-                    mylist.append(("I am executed first", sum))
+                    mylist.append("I am executed first")
     except Exception, e:
         print mylist[0]
         print e.args, sum
-'''
 
 def test_parallel_exceptions_unnested():
     """
@@ -453,7 +445,6 @@ def test_parallel_exceptions_unnested():
         print mylist[0]
         print e.args, sum
 
-'''
 cdef int parallel_exc_cdef() except -3:
     cdef int i, j
     for i in prange(10, nogil=True):
@@ -462,7 +453,6 @@ cdef int parallel_exc_cdef() except -3:
                 raise Exception("propagate me")
 
     return 0
-'''
 
 cdef int parallel_exc_cdef_unnested() except -3:
     cdef int i
@@ -480,9 +470,8 @@ def test_parallel_exc_cdef():
     Exception: propagate me
     """
     parallel_exc_cdef_unnested()
-    #parallel_exc_cdef()
+    parallel_exc_cdef()
 
-'''
 cpdef int parallel_exc_cpdef() except -3:
     cdef int i, j
     for i in prange(10, nogil=True):
@@ -491,7 +480,6 @@ cpdef int parallel_exc_cpdef() except -3:
                 raise Exception("propagate me")
 
     return 0
-'''
 
 cpdef int parallel_exc_cpdef_unnested() except -3:
     cdef int i, j
@@ -510,9 +498,8 @@ def test_parallel_exc_cpdef():
     Exception: propagate me
     """
     parallel_exc_cpdef_unnested()
-    #parallel_exc_cpdef()
+    parallel_exc_cpdef()
 
-'''
 cdef int parallel_exc_nogil_swallow() except -1:
     cdef int i, j
     for i in prange(10, nogil=True):
@@ -524,7 +511,6 @@ cdef int parallel_exc_nogil_swallow() except -1:
             return i
 
     return 0
-'''
 
 cdef int parallel_exc_nogil_swallow_unnested() except -1:
     cdef int i
@@ -542,13 +528,13 @@ def test_parallel_exc_nogil_swallow():
     """
     >>> test_parallel_exc_nogil_swallow()
     execute me
+    execute me
     """
     parallel_exc_nogil_swallow_unnested()
     print 'execute me'
-    #parallel_exc_nogil_swallow()
-    #print 'execute me'
+    parallel_exc_nogil_swallow()
+    print 'execute me'
 
-'''
 def parallel_exc_replace():
     """
     >>> parallel_exc_replace()
@@ -569,7 +555,13 @@ def parallel_exc_replace():
     return 0
 
 
-def _parallel_exceptions2():
+def parallel_exceptions2():
+    """
+    >>> parallel_exceptions2()
+    Traceback (most recent call last):
+       ...
+    Exception: propagate me
+    """
     cdef int i, j, k
 
     for i in prange(10, nogil=True):
@@ -582,59 +574,6 @@ def _parallel_exceptions2():
                     continue
                     return
 
-def test_parallel_exceptions2():
-    """
-    DISABLED
-
-    test_parallel_exceptions2()
-    read: start
-    propagate me
-    exiting...
-    Exit status: 0
-    """
-    if not hasattr(os, 'fork'):
-        print 'start'
-        print 'propagate me'
-        print 'Exit status: 0'
-        return
-
-    r, w = os.pipe()
-    fr = os.fdopen(r, 'r')
-    fw = os.fdopen(w, 'w', 0)
-
-    pid = os.fork()
-    if pid == 0:
-        try:
-            fr.close()
-            os.dup2(w, 1)
-            os.dup2(w, 2)
-
-            print >>fw, 'start'
-
-            try:
-                _parallel_exceptions2()
-            except Exception, e:
-                print >>fw, e.args[0]
-            else:
-                print >>fw, 'No exception caught'
-        except:
-            import traceback
-            print >>fw, traceback.format_exc()
-        finally:
-            print >>fw, 'exiting...'
-            os._exit(0)
-    else:
-        fw.close()
-
-        print 'read:', fr.read(),
-
-        pid, status = os.waitpid(pid, 0)
-        if os.WIFSIGNALED(status):
-            print 'Got signal', os.WTERMSIG(status)
-
-        print 'Exit status:', os.WEXITSTATUS(status)
-'''
-
 def test_parallel_with_gil_return():
     """
     >>> test_parallel_with_gil_return()
@@ -654,25 +593,6 @@ def test_parallel_with_gil_return():
         with gil:
             return sum
 
-'''
-def test_parallel_with_gil_continue():
-    """
-    >>> test_parallel_with_gil_continue()
-    20
-    """
-    cdef int i, sum = 0
-
-    for i in prange(10, nogil=True):
-        with cython.parallel.parallel():
-            with gil:
-                if i % 2:
-                    continue
-
-        sum += i
-
-    print sum
-'''
-
 def test_parallel_with_gil_continue_unnested():
     """
     >>> test_parallel_with_gil_continue_unnested()