Disable OpenMP nested parallelism

Cython/Compiler/Nodes.py

@@ -5910,6 +5910,22 @@ class ParallelStatNode(StatNode, ParallelNode):
         analyse_expressions phase
         """
         for entry, (pos, op) in self.assignments.iteritems():
+
+            if self.is_prange and not self.is_parallel:
+                # closely nested prange in a with parallel block, disallow
+                # assigning to privates in the with parallel block (we
+                # consider it too implicit and magicky for users)
+                if entry in self.parent.assignments:
+                    error(pos,
+                          "Cannot assign to private of outer parallel block, "
+                          "as we cannot retain its value after the loop")
+                    continue
+
+            if not self.is_prange and op:
+                # Again possible, but considered to magicky
+                error(pos, "Reductions not allowed for parallel blocks")
+                continue
+
             if self.is_private(entry):
                 # lastprivate = self.is_prange and entry == self.target.entry
                 # By default all variables should have the same values as if
@@ -6590,7 +6606,6 @@ class ParallelRangeNode(ParallelStatNode):
         if hasattr(self.target, 'entry'):
             self.assignments[self.target.entry] = self.target.pos, None
 
-        self.analyse_sharing_attributes(env)
         super(ParallelRangeNode, self).analyse_expressions(env)
 
     def nogil_check(self, env):
@@ -6723,10 +6738,12 @@ class ParallelRangeNode(ParallelStatNode):
 
         if not self.is_parallel:
             code.put("#pragma omp for")
+            self.privatization_insertion_point = code.insertion_point()
+            reduction_codepoint = self.parent.privatization_insertion_point
         else:
             code.put("#pragma omp parallel")
-            self.put_num_threads(code)
             self.privatization_insertion_point = code.insertion_point()
+            reduction_codepoint = self.privatization_insertion_point
             code.putln("")
             code.putln("#endif /* _OPENMP */")
 
@@ -6744,7 +6761,10 @@ class ParallelRangeNode(ParallelStatNode):
                 if entry.type.is_pyobject:
                     error(self.pos, "Python objects cannot be reductions")
                 else:
-                    code.put(" reduction(%s:%s)" % (op, entry.cname))
+                    #code.put(" reduction(%s:%s)" % (op, entry.cname))
+                    # This is the only way reductions + nesting works in gcc4.5
+                    reduction_codepoint.put(
+                                " reduction(%s:%s)" % (op, entry.cname))
             else:
                 if entry == self.target.entry:
                     code.put(" firstprivate(%s)" % entry.cname)
@@ -6762,9 +6782,7 @@ class ParallelRangeNode(ParallelStatNode):
         if self.schedule:
             code.put(" schedule(%s)" % self.schedule)
 
-        if not self.is_parallel:
-            self.put_num_threads(self.parent.privatization_insertion_point)
-            self.privatization_insertion_point = code.insertion_point()
+        self.put_num_threads(reduction_codepoint)
 
         code.putln("")
         code.putln("#endif /* _OPENMP */")

Cython/Compiler/TypeInference.py

@@ -26,6 +26,8 @@ class MarkAssignments(CythonTransform):
     # tells us whether we're in a normal loop
     in_loop = False
 
+    parallel_errors = False
+
     def __init__(self, context):
         super(CythonTransform, self).__init__()
         self.context = context
@@ -176,17 +178,30 @@ class MarkAssignments(CythonTransform):
         else:
             node.parent = None
 
+        nested = False
         if node.is_prange:
             if not node.parent:
                 node.is_parallel = True
             else:
                 node.is_parallel = (node.parent.is_prange or not
                                     node.parent.is_parallel)
+                nested = node.parent.is_prange
         else:
             node.is_parallel = True
+            # Note: nested with parallel() blocks are handled by
+            # ParallelRangeTransform!
+            # nested = node.parent
+            nested = node.parent and node.parent.is_prange
 
         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")
+            self.parallel_errors = True
+
         if node.is_prange:
             child_attrs = node.child_attrs
             node.child_attrs = ['body', 'target', 'args']
@@ -200,6 +215,7 @@ class MarkAssignments(CythonTransform):
             self.visitchildren(node)
             self.parallel_block_stack.pop()
 
+        self.parallel_errors = False
         return node
 
     def visit_YieldExprNode(self, node):

tests/errors/e_cython_parallel.pyx

@@ -76,6 +76,33 @@ def f(x):
             with gil:
                 yield x
 
+# Disabled nesting:
+
+for i in prange(10, nogil=True):
+    for y in prange(10):
+        pass
+
+with nogil, cython.parallel.parallel():
+    for i in prange(10):
+        for i in prange(10):
+            pass
+
+
+# Assign to private from parallel block in prange:
+cdef int myprivate1, myprivate2
+
+with nogil, cython.parallel.parallel():
+    myprivate1 = 1
+    for i in prange(10):
+        myprivate1 = 3
+        myprivate2 = 4
+    myprivate2 = 2
+
+# Disallow parallel with block reductions:
+i = 0
+with nogil, cython.parallel.parallel():
+    i += 1
+
 _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
@@ -97,4 +124,9 @@ 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, as we cannot retain its value after the loop
+e_cython_parallel.pyx:98:19: Cannot assign to private of outer parallel block, as we cannot retain its value after the loop
+e_cython_parallel.pyx:104:6: Reductions not allowed for parallel blocks
 """

tests/run/parallel.pyx

@@ -44,6 +44,7 @@ def test_num_threads():
 
     return num_threads
 
+'''
 def test_parallel_catch():
     """
     >>> test_parallel_catch()
@@ -67,6 +68,7 @@ def test_parallel_catch():
 
     print len(exceptions) == num_threads
     assert len(exceptions) == num_threads, (len(exceptions), num_threads)
+'''
 
 
 OPENMP_PARALLEL = True

tests/run/sequential_parallel.pyx

@@ -46,6 +46,7 @@ def test_descending_prange():
 
     return sum
 
+'''
 def test_propagation():
     """
     >>> test_propagation()
@@ -65,6 +66,7 @@ def test_propagation():
                     sum2 += y
 
     return i, j, x, y, sum1, sum2
+'''
 
 def test_unsigned_operands():
     """
@@ -224,7 +226,7 @@ def test_nan_init():
     cdef unsigned long d2 = 10
     cdef long long e1 = 10
     cdef unsigned long long e2 = 10
-    
+
     cdef actually_long_t miss1 = 10
     cdef actually_short_t miss2 = 10
     cdef myint_t typedef1 = 10
@@ -255,6 +257,27 @@ def test_nan_init():
             or typedef1 == 10):
             errp[0] = 1
 
+    cdef int i
+    for i in prange(10, nogil=True):
+        # First, trick the error checking to make it believe these variables
+        # are initialized after this if
+
+        if mybool: # mybool is always false!
+            a1 = a2 = b1 = b2 = c1 = c2 = d1 = d2 = e1 = e2 = 0
+            f = g = h = 0.0
+            p = NULL
+            miss1 = miss2 = typedef1 = 0
+
+        if (a1 == 10 or a2 == 10 or
+            b1 == 10 or b2 == 10 or
+            c1 == 10 or c2 == 10 or
+            d1 == 10 or d2 == 10 or
+            e1 == 10 or e2 == 10 or
+            f == 10.0 or g == 10.0 or h == 10.0 or
+            p == <void *> 10 or miss1 == 10 or miss2 == 10
+            or typedef1 == 10):
+            errp[0] = 1
+
     if err:
         raise Exception("One of the values was not initialized to a maximum "
                         "or NaN value")
@@ -263,9 +286,6 @@ def test_nan_init():
     with nogil, cython.parallel.parallel():
         c1 = 16
 
-    if globals().get('OPENMP_PARALLEL'):
-        # We only reach here when we are included by parallel.pyx
-        assert c1 == 20, c1
 
 cdef void nogil_print(char *s) with gil:
     print s.decode('ascii')
@@ -328,6 +348,7 @@ def test_prange_continue():
 
     free(p)
 
+'''
 def test_nested_break_continue():
     """
     >>> test_nested_break_continue()
@@ -358,6 +379,7 @@ def test_nested_break_continue():
                     continue
 
     print i
+'''
 
 cdef int parallel_return() nogil:
     cdef int i
@@ -377,6 +399,7 @@ def test_return():
     """
     print parallel_return()
 
+'''
 def test_parallel_exceptions():
     """
     >>> test_parallel_exceptions()
@@ -402,8 +425,34 @@ def test_parallel_exceptions():
     except Exception, e:
         print mylist[0]
         print e.args, sum
+'''
+
+def test_parallel_exceptions_unnested():
+    """
+    >>> test_parallel_exceptions_unnested()
+    ('I am executed first', 0)
+    ('propagate me',) 0
+    """
+    cdef int i, sum = 0
 
+    mylist = []
+
+    try:
+        with nogil, cython.parallel.parallel():
+            try:
+                for i in prange(10):
+                    with gil:
+                        raise Exception("propagate me")
+
+                    sum += i
+            finally:
+                with gil:
+                    mylist.append(("I am executed first", sum))
+    except Exception, e:
+        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):
@@ -412,6 +461,15 @@ cdef int parallel_exc_cdef() except -3:
                 raise Exception("propagate me")
 
     return 0
+'''
+
+cdef int parallel_exc_cdef_unnested() except -3:
+    cdef int i
+    for i in prange(10, nogil=True):
+        with gil:
+            raise Exception("propagate me")
+
+    return 0
 
 def test_parallel_exc_cdef():
     """
@@ -420,8 +478,10 @@ def test_parallel_exc_cdef():
         ...
     Exception: propagate me
     """
-    parallel_exc_cdef()
+    parallel_exc_cdef_unnested()
+    #parallel_exc_cdef()
 
+'''
 cpdef int parallel_exc_cpdef() except -3:
     cdef int i, j
     for i in prange(10, nogil=True):
@@ -430,6 +490,16 @@ 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
+    for i in prange(10, nogil=True):
+        with gil:
+            raise Exception("propagate me")
+
+    return 0
+
 
 def test_parallel_exc_cpdef():
     """
@@ -438,8 +508,10 @@ def test_parallel_exc_cpdef():
         ...
     Exception: propagate me
     """
-    parallel_exc_cpdef()
+    parallel_exc_cpdef_unnested()
+    #parallel_exc_cpdef()
 
+'''
 cdef int parallel_exc_nogil_swallow() except -1:
     cdef int i, j
     for i in prange(10, nogil=True):
@@ -451,15 +523,31 @@ cdef int parallel_exc_nogil_swallow() except -1:
             return i
 
     return 0
+'''
+
+cdef int parallel_exc_nogil_swallow_unnested() except -1:
+    cdef int i
+    with nogil:
+        try:
+            for i in prange(10):
+                with gil:
+                    raise Exception("propagate me")
+        finally:
+            return i
+
+    return 0
 
 def test_parallel_exc_nogil_swallow():
     """
     >>> test_parallel_exc_nogil_swallow()
     execute me
     """
-    parallel_exc_nogil_swallow()
+    parallel_exc_nogil_swallow_unnested()
     print 'execute me'
+    #parallel_exc_nogil_swallow()
+    #print 'execute me'
 
+'''
 def parallel_exc_replace():
     """
     >>> parallel_exc_replace()
@@ -480,7 +568,6 @@ def parallel_exc_replace():
     return 0
 
 
-
 def _parallel_exceptions2():
     cdef int i, j, k
 
@@ -545,6 +632,7 @@ def test_parallel_exceptions2():
             print 'Got signal', os.WTERMSIG(status)
 
         print 'Exit status:', os.WEXITSTATUS(status)
+'''
 
 def test_parallel_with_gil_return():
     """
@@ -565,6 +653,7 @@ def test_parallel_with_gil_return():
         with gil:
             return sum
 
+'''
 def test_parallel_with_gil_continue():
     """
     >>> test_parallel_with_gil_continue()
@@ -581,3 +670,20 @@ def test_parallel_with_gil_continue():
         sum += i
 
     print sum
+'''
+
+def test_parallel_with_gil_continue_unnested():
+    """
+    >>> test_parallel_with_gil_continue_unnested()
+    20
+    """
+    cdef int i, sum = 0
+
+    for i in prange(10, nogil=True):
+        with gil:
+            if i % 2:
+                continue
+
+        sum += i
+
+    print sum