Fix IndexError in memlet propagation

[edopao]

During integration of dace v0.15.1 into icon4py, the following exception was raised in SDFG auto-optimize step:

../edopao/gt4py/src/gt4py/next/program_processors/runners/dace_iterator/__init__.py:233: in build_sdfg_from_itir
    sdfg = autoopt.auto_optimize(sdfg, device, symbols=symbols, use_gpu_storage=on_gpu)
../edopao/dace/dace/transformation/auto/auto_optimize.py:575: in auto_optimize
    l2ms = sdfg.apply_transformations_repeated((LoopToMap, RefineNestedAccess),
../edopao/dace/dace/sdfg/sdfg.py:2484: in apply_transformations_repeated
    results = pazz.apply_pass(self, {})
../edopao/dace/dace/transformation/passes/pattern_matching.py:253: in apply_pass
    return self._apply_pass(sdfg, pipeline_results, apply_once=False)
../edopao/dace/dace/transformation/passes/pattern_matching.py:211: in _apply_pass
    self._apply_and_validate(match, sdfg, start, pipeline_results, applied_transformations)
../edopao/dace/dace/transformation/passes/pattern_matching.py:168: in _apply_and_validate
    applied_transformations[type(match).__name__].append(match.apply(graph, tsdfg))
../edopao/dace/dace/transformation/interstate/sdfg_nesting.py:1101: in apply
    propagation.propagate_memlets_state(sdfg, state)
../edopao/dace/dace/sdfg/propagation.py:1202: in propagate_memlets_state
    propagate_memlets_sdfg(node.sdfg)
../edopao/dace/dace/sdfg/propagation.py:1162: in propagate_memlets_sdfg
    propagate_memlets_state(sdfg, state)
../edopao/dace/dace/sdfg/propagation.py:1205: in propagate_memlets_state
    propagate_memlets_nested_sdfg(sdfg, state, node)
../edopao/dace/dace/sdfg/propagation.py:1084: in propagate_memlets_nested_sdfg
    border_memlet.src_subset[i] = fallback_subset[i]
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 

self = Range (0:geofac_n2s_shape0_6730), key = 1

    def __getitem__(self, key):
>       return self.ranges.__getitem__(key)
E       IndexError: list index out of range

../edopao/dace/dace/subsets.py:624: IndexError

This error is probably the same as one of the issues reported in #1468. The testcase in this PR allows to reproduce this issue.

This PR is applying a workaround in dace/sdfg/propagation.py, to make the test pass. However, this does not seem to be the proper the solution. This code change is rather a placeholder to highlight where the issue originates.

[edopao]

@phschaad This PR mainly includes a test SDFG to reproduce the issue in memlet propagation. The code change I made is not the fix, but rather a placeholder to highlight where this issue originates. It seems like you have worked on this code, so you maybe have some idea how we can properly fix it?

[philip-paul-mueller]

@edopao
I am not sure if this fix would also work if we slice off a row, i.e. A[idx_X, :]. But, I think it should be relatively easy^{(R)} to implement a general fix that works for that case and also in multiple dimensions.

[edopao]

Not an issue in memlet propagation, the problem was in the test SDFG. The memlet that performed the actual slicing was going from _out to _inp, i.e. in the wrong direction.