https://gcc.gnu.org/bugzilla/show_bug.cgi?id=87665
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Status|NEW |ASSIGNED
Component|c++ |tree-optimization
Assignee|unassigned at gcc dot gnu.org |rguenth at gcc dot
gnu.org
Target Milestone|--- |8.3
--- Comment #10 from Richard Biener <rguenth at gcc dot gnu.org> ---
--param vect-max-version-for-alias-checks=0 also avoids the issue. So it's
/usr/include/c++/8/bits/stl_uninitialized.h:82:23: note: loop vectorized
/usr/include/c++/8/bits/stl_uninitialized.h:82:23: note: loop versioned for
vectorization because of possible aliasing
that breaks things. This is vectorizing of
<bb 14> [local count: 163894221]:
# __first$_M_current_62 = PHI <_100(82), _84(86)>
# __cur_12 = PHI <iftmp.6_113(82), __cur_85(86)>
MEM[(struct &)__cur_12] ={v} {CLOBBER};
MEM[(struct &)__cur_12] ={v} {CLOBBER};
_82 = MEM[(int * const &)__first$_M_current_62];
MEM[(int * &)__first$_M_current_62] = 0B;
MEM[(struct &)__cur_12] ={v} {CLOBBER};
MEM[(struct _Head_base *)__cur_12]._M_head_impl = _82;
_83 = __first$_M_current_62->second;
__cur_12->second = _83;
_84 = __first$_M_current_62 + 16;
__cur_85 = __cur_12 + 16;
if (_67 != _84)
goto <bb 86>; [89.00%]
else
goto <bb 80>; [11.00%]
<bb 86> [local count: 145865857]:
goto <bb 14>; [100.00%]
where we end up interchanging the read and write from/to MEM[(int * const
&)__first$_M_current_62]. There are SLP and non-SLP stmts here,
the SLP ones being
stl_uninitialized.h:82:23: note: Final SLP tree for instance:
stl_uninitialized.h:82:23: note: node
stl_uninitialized.h:82:23: note: stmt 0 MEM[(struct _Head_base
*)__cur_12]._M_head_impl = _82;
stl_uninitialized.h:82:23: note: stmt 1 __cur_12->second = _83;
stl_uninitialized.h:82:23: note: node
stl_uninitialized.h:82:23: note: stmt 0 _82 = MEM[(int * const
&)__first$_M_current_62];
stl_uninitialized.h:82:23: note: stmt 1 _83 =
__first$_M_current_62->second;
