Suspected memory corruption in test_transform_normalize_array_shape_and_access

[reuterbal]

On MacOS, the following test fails with a variety of errors:

loki/loki/transformations/tests/test_array_indexing.py

Lines 323 to 439 in 2b3b206

	def test_transform_normalize_array_shape_and_access(tmp_path, frontend, start_index):
	"""
	Test normalization of array shape and access, thus changing arrays with start
	index different than "1" to have start index "1".
	E.g., ``x1(5:len)`` -> ```x1(1:len-4)``
	"""
	fcode = f"""
	module norm_arr_shape_access_mod
	implicit none
	contains
	subroutine norm_arr_shape_access(x1, x2, x3, x4, assumed_x1, l1, l2, l3, l4)
	! use nested_routine_mod, only : nested_routine
	implicit none
	integer :: i1, i2, i3, i4, c1, c2, c3, c4
	integer, intent(in) :: l1, l2, l3, l4
	integer, intent(inout) :: x1({start_index}:l1+{start_index}-1)
	integer, intent(inout) :: x2({start_index}:l2+{start_index}-1, &
	& {start_index}:l1+{start_index}-1)
	integer, intent(inout) :: x3({start_index}:l3+{start_index}-1, &
	& {start_index}:l2+{start_index}-1, {start_index}:l1+{start_index}-1)
	integer, intent(inout) :: x4({start_index}:l4+{start_index}-1, &
	& {start_index}:l3+{start_index}-1, {start_index}:l2+{start_index}-1, &
	& {start_index}:l1+{start_index}-1)
	integer, intent(inout) :: assumed_x1(l1)
	c1 = 1
	c2 = 1
	c3 = 1
	c4 = 1
	do i1=1,l1
	assumed_x1(i1) = c1
	call nested_routine(assumed_x1, l1, c1)
	end do
	x1({start_index}:l4+{start_index}-1) = 0
	do i1={start_index},l1+{start_index}-1
	x1(i1) = c1
	do i2={start_index},l2+{start_index}-1
	x2(i2, i1) = c2*10 + c1
	do i3={start_index},l3+{start_index}-1
	x3(i3, i2, i1) = c3*100 + c2*10 + c1
	do i4={start_index},l4+{start_index}-1
	x4(i4, i3, i2, i1) = c4*1000 + c3*100 + c2*10 + c1
	c4 = c4 + 1
	end do
	c3 = c3 + 1
	end do
	c2 = c2 + 1
	end do
	c1 = c1 + 1
	end do
	end subroutine norm_arr_shape_access
	subroutine nested_routine(nested_x1, l1, c1)
	implicit none
	integer, intent(in) :: l1, c1
	integer, intent(inout) :: nested_x1(:)
	integer :: i1
	do i1=1,l1
	nested_x1(i1) = c1
	end do
	end subroutine nested_routine
	end module norm_arr_shape_access_mod
	"""
	def init_arguments(l1, l2, l3, l4):
	x1 = np.zeros(shape=(l1,), order='F', dtype=np.int32)
	assumed_x1 = np.zeros(shape=(l1,), order='F', dtype=np.int32)
	x2 = np.zeros(shape=(l2,l1,), order='F', dtype=np.int32)
	x3 = np.zeros(shape=(l3,l2,l1,), order='F', dtype=np.int32)
	x4 = np.zeros(shape=(l4,l3,l2,l1,), order='F', dtype=np.int32)
	return x1, x2, x3, x4, assumed_x1
	def validate_routine(routine):
	arrays = [var for var in FindVariables().visit(routine.body) if isinstance(var, sym.Array)]
	for arr in arrays:
	assert all(not isinstance(shape, sym.RangeIndex) for shape in arr.shape)
	l1 = 2
	l2 = 3
	l3 = 4
	l4 = 5
	module = Module.from_source(fcode, frontend=frontend, xmods=[tmp_path])
	for routine in module.routines:
	normalize_range_indexing(routine) # Fix OMNI nonsense
	filepath = tmp_path/(f'norm_arr_shape_access_{frontend}.f90')
	# compile and test "original" module/function
	mod = jit_compile(module, filepath=filepath, objname='norm_arr_shape_access_mod')
	function = getattr(mod, 'norm_arr_shape_access')
	orig_x1, orig_x2, orig_x3, orig_x4, orig_assumed_x1 = init_arguments(l1, l2, l3, l4)
	function(orig_x1, orig_x2, orig_x3, orig_x4, orig_assumed_x1, l1, l2, l3, l4)
	clean_test(filepath)
	# apply `normalize_array_shape_and_access`
	for routine in module.routines:
	normalize_array_shape_and_access(routine)
	filepath = tmp_path/(f'norm_arr_shape_access_normalized_{frontend}.f90')
	# compile and test "normalized" module/function
	mod = jit_compile(module, filepath=filepath, objname='norm_arr_shape_access_mod')
	function = getattr(mod, 'norm_arr_shape_access')
	x1, x2, x3, x4, assumed_x1 = init_arguments(l1, l2, l3, l4)
	function(x1, x2, x3, x4, assumed_x1, l1, l2, l3, l4)
	clean_test(filepath)
	# validate the routine "norm_arr_shape_access"
	validate_routine(module.subroutines[0])
	# validate the nested routine to see whether the assumed size array got correctly handled
	assert module.subroutines[1].variable_map['nested_x1'] == 'nested_x1(:)'
	# check whether results generated by the "original" and "normalized" version agree
	assert (x1 == orig_x1).all()
	assert (assumed_x1 == orig_assumed_x1).all()
	assert (x2 == orig_x2).all()
	assert (x3 == orig_x3).all()
	assert (x4 == orig_x4).all()

The symptom is usually one of the following:

1. Pytest/Python crashes completely (sometimes belatedly) with a SIGABRT, backtrace suggests memory corruption (triggers in malloc sanity checks)
2. Wrong data type in the output of the transformed code:

        # check whether results generated by the "original" and "normalized" version agree
>       assert (x1 == orig_x1).all()
E       AttributeError: 'bool' object has no attribute 'all'

3. Output not matching expected results:

>       assert (x1 == orig_x1).all()
E       assert False
E        +  where False = <built-in method all of numpy.ndarray object at 0x11f9643f0>()
E        +    where <built-in method all of numpy.ndarray object at 0x11f9643f0> = array([1, 2], dtype=int32) == array([0, 2], dtype=int32)
E             
E             Full diff:
E             - array([0, 2], dtype=int32)
E             ?        ^
E             + array([1, 2], dtype=int32)
E             ?        ^.all