https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91825
Jason Merrill <jason at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Component|bootstrap |tree-optimization
Summary|Top-of-tree GCC does not |[10 regression] Bogus
|bootstrap (uninitialized |-Wmaybe-uninitialized with
|warning) |r275744 breaks bootstrap
--- Comment #7 from Jason Merrill <jason at gcc dot gnu.org> ---
OK, I'm confident now that this is not a problem with the front-end output.
The only difference from r275744 is that some indexing expressions are
represented in gimple as ar[idx] rather than *(&ar+idx), e.g.
_119 = scalar_int_mode::operator machine_mode
(&int_mode_to);
- _120 = (sizetype) _119;
- _121 = &mode_class + _120;
- _122 = *_121;
- if (_122 == 2) goto <D>; else goto <D>;
+ _120 = (long int) _119;
+ _121 = mode_class[_120];
+ if (_121 == 2) goto <D>; else goto <D>;
By the time we get to the uninit pass I see some differences in the gimple,
such as references to 'cstore'
+ # cstore_2229 = PHI <mode_69(90), mode_69(85), int_mode_87(D)(82),
mode_69(93
), mode_69(113)>
...
+ # cstore_2844 = PHI <mode_69(47), cstore_2229(100)>
...
- mode.96_671 = (int) mode_69;
+ mode.96_671 = (int) cstore_2844;
It's not clear to me how this affects uninit.
I can wait to reapply r275744 until I'm ready to commit the operator<=> patch,
but it would be good if someone could fix the tree-ssa issue before then.
