#2878 Add some support for dependency analysis involving array ranges.

Author: hiker

src/psyclone/psyir/tools/dependency_tools.py

@@ -48,7 +48,7 @@
 from psyclone.errors import InternalError, LazyString
 from psyclone.psyir.backend.sympy_writer import SymPyWriter
 from psyclone.psyir.backend.visitor import VisitorError
-from psyclone.psyir.nodes import Loop
+from psyclone.psyir.nodes import Loop, Node, Range
 
 
 # pylint: disable=too-many-lines
@@ -281,6 +281,50 @@ def _partition(comp_ind1, comp_ind2, loop_variables):
 
         return partition_infos
 
+    # -------------------------------------------------------------------------
+    @staticmethod
+    def _ranges_overlap(range1: Node,
+                        range2: Node) -> bool:
+        '''This function tests if two ranges overlap. It also accepts a simple
+        index as 'range' (e.g. just `i`), which will be converted into `i:i:1`
+        before comparing. At this stage, this function simple checks if one of
+        the ranges starts after the other (e.g. 1:3, and 5:7). It will handle
+        unspecified ranges (":"), and will report an overlap.
+        Additional tests e.g. using the step value are not yet implemented
+        (e.g. 1:10:2 and 2:10:2 will not overlap, but this will not be
+        detected atm).
+
+        :param range1: The first range or expression.
+        :param range2: The second range or expression.
+
+        :returns: whether the ranges (or an index expression with a range)
+            overlap or not
+
+        '''
+        if not isinstance(range1, Range):
+            # Not a range, must be some index `i`. Create a range `i:i:1`
+            range1 = Range.create(range1.copy(), range1.copy())
+        if not isinstance(range2, Range):
+            # Not a range, must be some index `i`. Create a range `i:i:1`
+            range2 = Range.create(range2.copy(), range2.copy())
+
+        sm = SymbolicMaths.get()
+
+        # Check if the first range is smaller than the second one, e.g.:
+        # 1:3:1 and 4:6:1
+        if sm.greater_than(range2.start, range1.stop) == sm.Fuzzy.TRUE:
+            # The first range is before the second range, so no overlap
+            return False
+        # Check if the second range is smaller than the first one, e.g.:
+        # 4:6:1 and 1:3:1
+        if sm.greater_than(range1.start, range2.stop) == sm.Fuzzy.TRUE:
+            # The second range is before the first range, so no overlap
+            return False
+
+        # We could do additional tests here, e.g. including step to determine
+        # that 1:10:2 does not overlap with 2:10:2
+        return True
+
     # -------------------------------------------------------------------------
     @staticmethod
     def _independent_0_var(index_exp1, index_exp2):
@@ -296,6 +340,9 @@ def _independent_0_var(index_exp1, index_exp2):
         :type index_exp2: :py:class:`psyclone.psyir.nodes.Node`
 
         '''
+        if isinstance(index_exp1, Range) or isinstance(index_exp2, Range):
+            return not DependencyTools._ranges_overlap(index_exp1, index_exp2)
+
         sym_maths = SymbolicMaths.get()
 
         # If the indices can be shown to be never equal, the accesses

src/psyclone/tests/psyir/tools/dependency_tools_test.py

@@ -41,7 +41,7 @@
 from psyclone.configuration import Config
 from psyclone.core import AccessType, Signature, VariablesAccessInfo
 from psyclone.errors import InternalError
-from psyclone.psyir.nodes import Loop
+from psyclone.psyir.nodes import Assignment, Loop
 from psyclone.psyir.tools import DependencyTools, DTCode
 from psyclone.tests.utilities import get_invoke
 
@@ -1107,3 +1107,67 @@ def test_fuse_dimension_change(fortran_reader):
             "in different index locations: s%comp1(jj)%comp2(ji) and "
             "s%comp1(ji)%comp2(jj)."
             in str(msg))
+
+
+# ----------------------------------------------------------------------------
+@pytest.mark.parametrize("range1, range2, overlap",
+                         [("1:3", "4:6", False),
+                          ("3:9", "-1:-3", False),
+                          ("1:3", "4", False),
+                          ("5", "-1:-3", False),
+                          ("i:i+3", "i+5:i+7", False),
+                          ("i:i+3", "i+2", True),
+                          ("i:i+3", "i+5", False),
+                          ("i:i+3", "i-1", False),
+                          (":", "1", True),
+                          (":", "i", True),
+                          ("::", "1", True),
+                          ("::", "i", True),
+                          ("1", ":", True),
+                          ("i", ":", True),
+                          ])
+def test_ranges_overlap(range1, range2, overlap, fortran_reader):
+    '''Test the detection of overlapping ranges.
+    '''
+    source = f'''program test
+                 integer i, ji, inbj
+                 integer, parameter :: jpi=5, jpj=10
+                 real, dimension(jpi,jpi) :: ldisoce
+
+                 ldisoce({range1},{range2}) = 1.0
+                 end program test'''
+
+    psyir = fortran_reader.psyir_from_source(source)
+    dep_tools = DependencyTools()
+    assign = psyir.walk(Assignment)[0]
+    r1 = assign.lhs.children[0]
+    r2 = assign.lhs.children[1]
+    assert dep_tools._ranges_overlap(r1, r2) == overlap
+    # Also make sure that _independent_0_var handles this correctly:
+    assert dep_tools._independent_0_var(r1, r2) is not overlap
+
+
+# ----------------------------------------------------------------------------
+def test_nemo_example_ranges(fortran_reader):
+    '''Tests an actual NEMO example
+    '''
+    source = '''program test
+                integer ji, inbj
+                integer, parameter :: jpi=5, jpj=10
+                real, dimension(jpi,jpi) :: ldisoce
+                do jj = 1, inbj, 1
+                  if (COUNT(ldisoce(:,jj)) == 0) then
+                    ldisoce(1,jj) = .true.
+                  end if
+                enddo
+                end program test'''
+
+    psyir = fortran_reader.psyir_from_source(source)
+    loops = psyir.children[0].children[0]
+    dep_tools = DependencyTools()
+
+    # This loop can be parallelised because all instances of ldisoce use
+    # the index jj in position 2 (the overlap between ":" and "1"
+    # is tested in test_ranges_overlap above, here we check that this
+    # overlap is indeed ignored because of the jj index).
+    assert dep_tools.can_loop_be_parallelised(loops)