[AUTOMATED] Read DWARF 5 indexed variable locations in TypeMatch

decbench/metrics/type_match.py

@@ -528,8 +528,9 @@ def _get_location(die: Any, dwarfinfo: Any) -> tuple[list[int], bool]:
                 if op.op_name == "DW_OP_fbreg":
                     offsets.append(op.args[0] + 16)
 
-        elif loc_attr.form == "DW_FORM_sec_offset":
+        elif loc_attr.form in ("DW_FORM_sec_offset", "DW_FORM_loclistx"):
             loclists = dwarfinfo.location_lists()
+            # pyelftools has resolved a loclistx index to a section offset in value.
             loclist = loclists.get_location_list_at_offset(loc_attr.value, die=die)
 
             for entry in loclist:
@@ -841,7 +842,7 @@ class TypeMatchMetric(Metric):
     display_name = "Type Correctness"
     description = "Accuracy of variable type recovery vs DWARF ground truth"
 
-    cache_version = "18"
+    cache_version = "20"
 
     weight = 1.0
     lower_is_better = False

docs/metrics.md

@@ -118,8 +118,8 @@ Ground truth is the function's OWN parameters and locals: the variables of an
 inlined callee (`DW_TAG_inlined_subroutine`, at `-O2` mostly the parameters of
 glibc's `_FORTIFY_SOURCE` wrappers such as `printf`'s `__fmt`) are not scored,
 since the compiled function usually has no storage for them.
-Current `cache_version="18"`, bumped for scoped syntax-aware declaration parsing,
-aggregate/qualifier normalization, and DWARF restrict-type passthrough.
+Current `cache_version="20"`, bumped to include DWARF 5 indexed location lists
+(`DW_FORM_loclistx`) when extracting ground-truth variables.
 The per-function key covers source and
 decompiler address evidence, argument/stack anchors, recovered types, source
 selection diagnostics, and the matching thresholds.

tests/test_metrics.py

@@ -432,6 +432,34 @@ def __init__(self, index: int, *, entry: bool, exit: bool):
 class TestTypeMatchMetric:
     """Tests for the type match metric."""
 
+    @pytest.mark.parametrize("form", ["DW_FORM_sec_offset", "DW_FORM_loclistx"])
+    @pytest.mark.parametrize(
+        ("expression", "expected"),
+        [([0x50], ([], True)), ([0x91, 0x78], ([8], True)), ([], ([], False))],
+    )
+    def test_location_lists_use_decoded_offset(
+        self, form: str, expression: list[int], expected: tuple[list[int], bool]
+    ) -> None:
+        from types import SimpleNamespace
+
+        from elftools.dwarf.structs import DWARFStructs
+
+        from decbench.metrics.type_match import _get_location
+
+        attribute = SimpleNamespace(form=form, value=0x38, raw_value=2)
+        die = SimpleNamespace(attributes={"DW_AT_location": attribute})
+
+        def read_list(offset: int, *, die: SimpleNamespace) -> list[SimpleNamespace]:
+            assert offset == 0x38
+            assert die.attributes["DW_AT_location"] is attribute
+            return [SimpleNamespace(loc_expr=expression)]
+
+        dwarf = SimpleNamespace(
+            structs=DWARFStructs(little_endian=True, dwarf_format=32, address_size=8),
+            location_lists=lambda: SimpleNamespace(get_location_list_at_offset=read_list),
+        )
+        assert _get_location(die, dwarf) == expected
+
     def test_type_match_registration(self) -> None:
         MetricRegistry.clear()
         from decbench.metrics.type_match import TypeMatchMetric