OpenMP Control Flow

Cython/Compiler/FlowControl.py

@@ -700,7 +700,9 @@ class CreateControlFlowGraph(CythonTransform):
         return node
 
     def visit_InPlaceAssignmentNode(self, node):
+        self.in_inplace_assignment = True
         self.visitchildren(node)
+        self.in_inplace_assignment = False
         self.mark_assignment(node.lhs, node.create_binop_node())
         return node
 
@@ -726,6 +728,11 @@ class CreateControlFlowGraph(CythonTransform):
             entry = node.entry or self.env.lookup(node.name)
             if entry:
                 self.flow.mark_reference(node, entry)
+
+                if entry in self.reductions and not self.in_inplace_assignment:
+                    error(node.pos,
+                          "Cannot read reduction variable in loop body")
+
         return node
 
     def visit_StatListNode(self, node):
@@ -806,10 +813,16 @@ class CreateControlFlowGraph(CythonTransform):
         # Target assignment
         self.flow.nextblock()
         self.mark_assignment(node.target)
+
         # Body block
+        if isinstance(node, Nodes.ParallelRangeNode):
+            # In case of an invalid
+            self._delete_privates(node, exclude=node.target.entry)
+
         self.flow.nextblock()
         self.visit(node.body)
         self.flow.loops.pop()
+
         # Loop it
         if self.flow.block:
             self.flow.block.add_child(condition_block)
@@ -828,11 +841,38 @@ class CreateControlFlowGraph(CythonTransform):
             self.flow.block = None
         return node
 
+    def _delete_privates(self, node, exclude=None):
+        for private_node in node.assigned_nodes:
+            if not exclude or private_node.entry is not exclude:
+                self.flow.mark_deletion(private_node, private_node.entry)
+
     def visit_ParallelRangeNode(self, node):
-        # if node.target is None an error will have been previously issued
-        if node.target is not None:
+        reductions = self.reductions
+
+        # if node.target is None or not a NameNode, an error will have
+        # been previously issued
+        if hasattr(node.target, 'entry'):
+            self.reductions = cython.set(reductions)
+
+            for private_node in node.assigned_nodes:
+                private_node.entry.error_on_uninitialized = True
+                pos, reduction = node.assignments[private_node.entry]
+                if reduction:
+                    self.reductions.add(private_node.entry)
+
             node = self.visit_ForInStatNode(node)
 
+        self.reductions = reductions
+        return node
+
+    def visit_ParallelWithBlockNode(self, node):
+        for private_node in node.assigned_nodes:
+            private_node.entry.error_on_uninitialized = True
+
+        self._delete_privates(node)
+        self.visitchildren(node)
+        self._delete_privates(node)
+
         return node
 
     def visit_ForFromStatNode(self, node):

Cython/Compiler/Nodes.py

@@ -5884,6 +5884,9 @@ class ParallelStatNode(StatNode, ParallelNode):
         # If op is not None, it's a reduction.
         self.privates = {}
 
+        # [NameNode]
+        self.assigned_nodes = []
+
     def analyse_declarations(self, env):
         self.body.analyse_declarations(env)
 
@@ -5901,7 +5904,6 @@ class ParallelStatNode(StatNode, ParallelNode):
     def analyse_expressions(self, env):
         self.body.analyse_expressions(env)
         self.analyse_sharing_attributes(env)
-        self.check_independent_iterations()
 
     def analyse_sharing_attributes(self, env):
         """
@@ -5917,8 +5919,7 @@ class ParallelStatNode(StatNode, ParallelNode):
                 # 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")
+                          "Cannot assign to private of outer parallel block")
                     continue
 
             if not self.is_prange and op:
@@ -5926,20 +5927,10 @@ class ParallelStatNode(StatNode, ParallelNode):
                 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
-                # executed sequentially
-                lastprivate = True
-                self.propagate_var_privatization(entry, op, lastprivate)
-
-    def is_private(self, entry):
-        """
-        True if this scope should declare the variable private, lastprivate
-        or reduction.
-        """
-        return (self.is_parallel or
-                (self.parent and entry not in self.parent.privates))
+            # By default all variables should have the same values as if
+            # executed sequentially
+            lastprivate = True
+            self.propagate_var_privatization(entry, op, lastprivate)
 
     def propagate_var_privatization(self, entry, op, lastprivate):
         """
@@ -6029,40 +6020,6 @@ class ParallelStatNode(StatNode, ParallelNode):
         code.putln("%s = %s;" % (cname, entry.cname))
         entry.cname = cname
 
-    def check_independent_iterations(self):
-        """
-        This checks for uninitialized thread-private variables, it's far from
-        fool-proof as it does not take control flow into account, nor
-        assignment to a variable as both the lhs and rhs. So it detects only
-        cases like this:
-
-            for i in prange(10, nogil=True):
-                var = x # error, x is private and read before assigned
-                x = i
-
-        Fortunately, it doesn't need to be perfect, as we still initialize
-        private variables to "invalid" values, such as NULL or NaN whenever
-        possible.
-        """
-        from Cython.Compiler import ParseTreeTransforms
-
-        transform = ParseTreeTransforms.FindUninitializedParallelVars()
-        transform(self.body)
-
-        for entry, pos in transform.used_vars:
-            if entry in self.privates:
-                assignment_pos, op = self.assignments[entry]
-
-                # Reading reduction variables is valid (in fact necessary)
-                # before assignment
-                if not op and pos < assignment_pos:
-                    if self.is_prange:
-                        error(pos, "Expression value depends on previous loop "
-                                   "iteration, cannot execute in parallel")
-                    else:
-                        error(pos, "Expression depends on an uninitialized "
-                                   "thread-private variable")
-
     def initialize_privates_to_nan(self, code, exclude=None):
         first = True
 
@@ -6739,11 +6696,11 @@ 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
+            # reduction_codepoint = self.parent.privatization_insertion_point
         else:
             code.put("#pragma omp parallel")
             self.privatization_insertion_point = code.insertion_point()
-            reduction_codepoint = self.privatization_insertion_point
+            # reduction_codepoint = self.privatization_insertion_point
             code.putln("")
             code.putln("#endif /* _OPENMP */")
 
@@ -6755,6 +6712,11 @@ class ParallelRangeNode(ParallelStatNode):
             code.putln("#ifdef _OPENMP")
             code.put("#pragma omp for")
 
+        # Nested parallelism is not supported, so we can put reductions on the
+        # for and not on the parallel (but would be valid, but gcc45 bugs on
+        # the former)
+        reduction_codepoint = code
+
         for entry, (op, lastprivate) in self.privates.iteritems():
             # Don't declare the index variable as a reduction
             if op and op in "+*-&^|" and entry != self.target.entry:

Cython/Compiler/ParseTreeTransforms.py

@@ -2338,24 +2338,6 @@ class TransformBuiltinMethods(EnvTransform):
         return node
 
 
-class FindUninitializedParallelVars(CythonTransform, SkipDeclarations):
-    """
-    This transform isn't part of the pipeline, it simply finds all references
-    to variables in parallel blocks.
-    """
-
-    def __init__(self):
-        CythonTransform.__init__(self, None)
-        self.used_vars = []
-
-    def visit_ParallelStatNode(self, node):
-        return node
-
-    def visit_NameNode(self, node):
-        self.used_vars.append((node.entry, node.pos))
-        return node
-
-
 class DebugTransform(CythonTransform):
     """
     Write debug information for this Cython module.

Cython/Compiler/Pipeline.py

@@ -148,9 +148,10 @@ def create_pipeline(context, mode, exclude_classes=()):
         EmbedSignature(context),
         EarlyReplaceBuiltinCalls(context),  ## Necessary?
         TransformBuiltinMethods(context),  ## Necessary?
+        MarkAssignments(context),
         CreateControlFlowGraph(context),
         RemoveUnreachableCode(context),
-        MarkAssignments(context),
+        # MarkAssignments(context),
         MarkOverflowingArithmetic(context),
         IntroduceBufferAuxiliaryVars(context),
         _check_c_declarations,

Cython/Compiler/TypeInference.py

@@ -62,6 +62,7 @@ class MarkAssignments(CythonTransform):
                     pos = lhs.pos
 
                 parallel_node.assignments[lhs.entry] = (pos, inplace_op)
+                parallel_node.assigned_nodes.append(lhs)
 
         elif isinstance(lhs, ExprNodes.SequenceNode):
             for arg in lhs.args:

tests/errors/e_cython_parallel.pyx

@@ -103,6 +103,22 @@ i = 0
 with nogil, cython.parallel.parallel():
     i += 1
 
+# Use of privates after the parallel with block
+with nogil, cython.parallel.parallel():
+    i = 1
+
+print i
+i = 2
+print i
+
+# Reading of reduction variables in the prange block
+cdef int sum = 0
+for i in prange(10, nogil=True):
+    sum += i
+    with gil:
+        print sum
+
+
 _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
@@ -117,8 +133,8 @@ 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:39:12: The parallel directive must be called
-e_cython_parallel.pyx:45:10: Expression value depends on previous loop iteration, cannot execute in parallel
-e_cython_parallel.pyx:55:9: Expression depends on an uninitialized thread-private variable
+e_cython_parallel.pyx:45:10: local variable 'y' referenced before assignment
+e_cython_parallel.pyx:55:9: local variable 'y' referenced before assignment
 e_cython_parallel.pyx:60:6: Reduction operator '*' is inconsistent with previous reduction operator '+'
 e_cython_parallel.pyx:62:36: cython.parallel.parallel() does not take positional arguments
 e_cython_parallel.pyx:65:36: Invalid keyword argument: invalid
@@ -126,7 +142,9 @@ 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: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
+e_cython_parallel.pyx:110:7: local variable 'i' referenced before assignment
+e_cython_parallel.pyx:119:17: Cannot read reduction variable in loop body
 """

tests/run/sequential_parallel.pyx

@@ -79,10 +79,11 @@ def test_unsigned_operands():
     cdef int step = 1
 
     cdef int steps_taken = 0
+    cdef int *steps_takenp = &steps_taken
 
     for i in prange(start, stop, step, nogil=True):
         steps_taken += 1
-        if steps_taken > 10:
+        if steps_takenp[0] > 10:
             abort()
 
     return steps_taken
@@ -168,7 +169,7 @@ def test_closure_parallel_with_gil():
         for i in prange(10, nogil=True):
             with gil:
                 sum += temp1 + temp2 + i
-                assert abs(sum - sum) == 0
+                # assert abs(sum - sum) == 0
 
         return sum