[DWARFLinker][Determinism] Clear the string offsets map of a unit that is cloned again (#240432)
While trying to optimize the parallel linker, I found that the output
wasn't always deterministic. Tracking it down with an agent, I found
this edge case where maybeResetToLoadedStage() doesn't clear the
.debug_str_offsets.
Context from agent:
When the parallel linker finds that an already cloned compile unit is
referenced by another unit, it resets the unit with
maybeResetToLoadedStage() and clones it again. The reset cleared the
abbreviations, the address index map and the output sections, but not
the map from strings to their .debug_str_offsets indexes. The second
cloning therefore kept the indexes assigned by the first one, and those
depend on whether, and how far, the unit had been cloned when it was
found to be referenced. With DWARF 5 inputs, the DW_FORM_strx indexes,
the .debug_str_offsets contents and the order of .debug_str could change
from one run to the next.
Clear the map with the rest of the per-cloning state.
The new test links a unit that is referenced by a later unit, so it is
cloned twice with one thread, and checks that its string indexes match a
single cloning.
Assisted-by: Claude