http://gcc.gnu.org/bugzilla/show_bug.cgi?id=22488
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Target Milestone|4.6.4 |---
Summary|[4.6/4.7/4.8/4.9 |C++ frontend generates
|Regression] |RECORD_TYPEs with
|push_fields_onto_fieldstack |overlapping FIELD_DECLs
|calculates offset |
|incorrectly |
--- Comment #57 from Richard Biener <rguenth at gcc dot gnu.org> 2013-04-02
08:04:05 UTC ---
Re-confirmed and summary adjusted. The code in tree-ssa-structalias.c is
still there:
/* We can't sort them if we have a field with a variable sized type,
which will make notokay = true. In that case, we are going to return
without creating varinfos for the fields anyway, so sorting them is a
waste to boot. */
if (!notokay)
{
sort_fieldstack (fieldstack);
/* Due to some C++ FE issues, like PR 22488, we might end up
what appear to be overlapping fields even though they,
in reality, do not overlap. Until the C++ FE is fixed,
we will simply disable field-sensitivity for these cases. */
notokay = check_for_overlaps (fieldstack);
I am not aware of other code in the middle-end that might be confused
about this issue (well, SRA maybe, but it handles overlaps well as it
also handles accesses to unions).
I believe that the C++ frontend issue might not be a regression after all,
the tree-ssa-structalias.c "regression" was fixed by the above kludge.
The C++ frontend issue makes it impossible to implement a GENERIC verifier
for RECORD_TYPEs (verify FIELD_DECLs are sorted after offset and not
overlapping).
