https://gcc.gnu.org/bugzilla/show_bug.cgi?id=78158
--- Comment #4 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
That is clearly due to bugs in tsan.c, in my PR55439 r194133 changes.
At -O0 the code is not optimized, so we end up with e.g.
__m_6 = 5;
__b_7 = std::operator& (__m_6, 65535);
__m.10_8 = (int) __m_6;
_9 = &this_5->_M_i;
_10 = __atomic_load_1 (_9, __m.10_8);
or
__m_12 = __m_7(D);
_13 = std::__cmpexch_failure_order (__m_12);
_14 = (int) __i2_11;
this_15 = this_9;
__i1_16 = __i1_10;
__i2_17 = (__int_type) _14;
__m1_18 = __m_12;
__m2_19 = _13;
__b2_20 = std::operator& (__m2_19, 65535);
__b1_21 = std::operator& (__m1_18, 65535);
__m2.16_22 = (int) __m2_19;
__m1.17_23 = (int) __m1_18;
_24 = (int) __i2_17;
_25 = &this_15->_M_i;
_26 = __atomic_compare_exchange_1 (_25, __i1_16, _24, 0, __m1.17_23,
__m2.16_22);
and then we run into instrument_builtin_call:
case check_last:
case fetch_op:
last_arg = gimple_call_arg (stmt, num - 1);
if (!tree_fits_uhwi_p (last_arg)
|| memmodel_base (tree_to_uhwi (last_arg)) >= MEMMODEL_LAST)
return; // <<<<======== HERE
...
case weak_cas:
if (!integer_nonzerop (gimple_call_arg (stmt, 3)))
continue; // <<<<======== HERE
/* FALLTHRU */
case strong_cas:
gcc_assert (num == 6);
for (j = 0; j < 6; j++)
args[j] = gimple_call_arg (stmt, j);
if (!tree_fits_uhwi_p (args[4])
|| memmodel_base (tree_to_uhwi (args[4])) >= MEMMODEL_LAST)
return; // <<<<======== HERE
if (!tree_fits_uhwi_p (args[5])
|| memmodel_base (tree_to_uhwi (args[5])) >= MEMMODEL_LAST)
return; // <<<<======== HERE
...
if (!tree_fits_uhwi_p (last_arg)
|| memmodel_base (tree_to_uhwi (last_arg)) >= MEMMODEL_LAST)
return; // <<<<======== HERE
If last_arg or gimple_call_arg (stmt, 3) is not an integer, but SSA_NAME, we
could for some very limited cases try to iterate over SSA_NAME definition
statements and see if we can figure out a constant, but as can be seen on the
snippets above, that is not really useful here, there are function calls which
at -O0 are not inlined.
Thus I think what we need to do is when we can't check the memory model value
at compile time, emit a runtime check which does the same thing, and use the
tsan builtins if the runtime test(s) pass, otherwise fall through into the
__atomic_* builtins.
