Fixing declaration of inner OpenMP privates (GH-3348)

* omp declare privates on outer loop, since inner loops are not 'omp parallel for'ified

Cython/Compiler/Nodes.py

@@ -9240,12 +9240,15 @@ class ParallelRangeNode(ParallelStatNode):
         code.putln("%(target)s = (%(target_type)s)(%(start)s + %(step)s * %(i)s);" % fmt_dict)
         self.initialize_privates_to_nan(code, exclude=self.target.entry)
 
-        if self.is_parallel:
+        if self.is_parallel and not self.is_nested_prange:
+            # nested pranges are not omp'ified, temps go to outer loops
             code.funcstate.start_collecting_temps()
 
         self.body.generate_execution_code(code)
         self.trap_parallel_exit(code, should_flush=True)
-        self.privatize_temps(code)
+        if self.is_parallel and not self.is_nested_prange:
+            # nested pranges are not omp'ified, temps go to outer loops
+            self.privatize_temps(code)
 
         if self.breaking_label_used:
             # Put a guard around the loop body in case return, break or

tests/run/sequential_parallel.pyx

@@ -378,11 +378,7 @@ def test_prange_continue():
 
 def test_nested_break_continue():
     """
-    DISABLED. For some reason this fails intermittently on jenkins, with
-    the first line of output being '0 0 0 0'. The generated code looks
-    awfully correct though... needs investigation
-
-    >> test_nested_break_continue()
+    >>> test_nested_break_continue()
     6 7 6 7
     8
     """
@@ -770,3 +766,53 @@ def test_prange_in_with(int x, ctx):
             with gil:
                 l[0] += i
         return l[0]
+
+
+cdef extern from *:
+    """
+    #ifdef _OPENMP
+    #define _get_addr(_x, _idx) &_x
+    #else
+    #define _get_addr(_x, _idx) (&_x+_idx)
+    #endif
+    #define address_of_temp(store, temp, idx) store = _get_addr(temp, idx)
+    #define address_of_temp2(store, ignore, temp, idx) store = _get_addr(temp, idx)
+
+    double get_value() {
+        return 1.0;
+    }
+    """
+    void address_of_temp(...) nogil
+    void address_of_temp2(...) nogil
+    double get_value() nogil except -1.0  # will generate a temp for exception checking
+
+def test_inner_private():
+    """
+    Determines if a temp variable is private by taking its address in multiple threads
+    and seeing if they're the same (thread private variables should have different
+    addresses
+    >>> test_inner_private()
+    ok
+    """
+    cdef double* not_parallel[2]
+    cdef double* inner_vals[2]
+    cdef double* outer_vals[2]
+    cdef Py_ssize_t n, m
+
+    for n in range(2):
+        address_of_temp(not_parallel[n], get_value(), 0)
+    assert not_parallel[0] == not_parallel[1], "Addresses should be the same since they come from the same temp"
+
+    for n in prange(2, num_threads=2, schedule='static', chunksize=1, nogil=True):
+        address_of_temp(outer_vals[n], get_value(), n)
+        for m in prange(1):
+            # second temp just ensures different numbering
+            address_of_temp2(inner_vals[n], get_value(), get_value(), n)
+
+    inner_are_the_same = inner_vals[0] == inner_vals[1]
+    outer_are_the_same = outer_vals[0] == outer_vals[1]
+
+    assert outer_are_the_same == False, "Temporary variables in outer loop should be private"
+    assert inner_are_the_same == False,  "Temporary variables in inner loop should be private"
+
+    print('ok')