https://gcc.gnu.org/bugzilla/show_bug.cgi?id=89579
Jakub Jelinek changed:
What|Removed |Added
CC||jakub at gcc dot gnu.org,
||law at gcc dot gnu.org,
||redi at gcc dot gnu.org
--- Comment #1 from Jakub Jelinek ---
This is caused by the uninitialized __capacity variable in
basic_string& operator=(basic_string&& __str);
The code is like:
// Just move the allocated pointer, our allocator can free it.
pointer __data = nullptr;
size_type __capacity;
if (!_M_is_local())
{
if (_Alloc_traits::_S_always_equal())
{
// __str can reuse our existing storage.
__data = _M_data();
__capacity = _M_allocated_capacity;
}
else // __str can't use it, so free it.
_M_destroy(_M_allocated_capacity);
}
_M_data(__str._M_data());
_M_length(__str.length());
_M_capacity(__str._M_allocated_capacity);
if (__data)
{
__str._M_data(__data);
__str._M_capacity(__capacity);
}
else
__str._M_data(__str._M_local_buf);
Although __capacity is never actually used uninitialized (and e.g. the uninit
warning pass figures that out), because it is used only if __data is non-NULL
and __data is non-NULL only if __capacity is also initialized, with -Og or e.g.
with -O1 -fno-tree-dominator-opts, if we don't perform jump threading, we end
up with something like:
if (_M_local_buf != _35)
goto ; [70.00%]
else
goto ; [30.00%]
[local count: 173485181]:
__capacity_44 = dummy.D.19123._M_allocated_capacity;
[local count: 247835973]:
# __data_47 = PHI <_35(14), 0B(13)>
# __capacity_48 = PHI <__capacity_44(14), __capacity_50(D)(13)>
dummy._M_dataplus._M_p = _37;
_45 = D.24766._M_string_length;
dummy._M_string_length = _45;
_46 = D.24766.D.19123._M_allocated_capacity;
dummy.D.19123._M_allocated_capacity = _46;
if (__data_47 != 0B)
goto ; [70.00%]
else
goto ; [30.00%]
[local count: 173485181]:
D.24766._M_dataplus._M_p = __data_47;
D.24766.D.19123._M_allocated_capacity = __capacity_48;
goto ; [100.00%]
before expansion and later when IRA checks for the clobbered vars, it does:
if (VAR_P (decl)
&& DECL_RTL_SET_P (decl)
&& REG_P (DECL_RTL (decl))
&& regno_clobbered_at_setjmp (setjmp_crosses, REGNO (DECL_RTL
(decl
where the last one uses:
return ((REG_N_SETS (regno) > 1
|| REGNO_REG_SET_P (df_get_live_out (ENTRY_BLOCK_PTR_FOR_FN (cfun)),
regno))
&& REGNO_REG_SET_P (setjmp_crosses, regno));
While REG_N_SETS is just 1 here for the pseudo used for __capacity, because of
the uninitialized use in the PHI it is considered live on the entry block and
that is why the warning is emitted.
Changing libstdc++ code to size_type __capacity = 0; makes the warning away,
though we'd need to check if it doesn't result in worse generated code.
