https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91004
--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> ---
Before the rev. DOM2 removes the call to __warn_memset_zero_len while with
the change the condition guarding the call changes like this:
<bb 3> [local count: 1073741824]:
_4 = MEM[(unsigned int *)this_2(D) + 12B];
- # RANGE [0, 4294967295] NONZERO 4294967295
- _6 = (long unsigned int) _4;
- _7 = __builtin_constant_p (_6);
- _11 = _6 == 0;
- _12 = _7 != 0;
- _13 = _11 & _12;
- if (_13 != 0)
+ if (_4 == 0)
goto <bb 4>; [100.00%]
else
goto <bb 5>; [0.00%]
<bb 4> [local count: 1073741824]:
# USE = nonlocal
# CLB = nonlocal
__warn_memset_zero_len ();
<bb 5> [count: 0]:
__builtin_unreachable ();
and the first change caused by the rev. is
--- a/pr34850.C.025t.einline 2019-06-27 10:53:11.462941555 +0200
+++ b/pr34850.C.025t.einline 2019-06-27 10:52:27.146152524 +0200
@@ -71,6 +71,7 @@
<bb 2> :
_1 = (long unsigned int) n_2(D);
+ # RANGE ~[0, 0]
_6 = __builtin_constant_p (_1);
if (_6 != 0)
goto <bb 3>; [100.00%]
else
goto <bb 5>; [0.00%]
<bb 5> [count: 0]:
__builtin_unreachable ();
which causes the __builtin_constant_p call to be eliminated way earlier
which IMHO is reasonable in principle but it breaks the
void * memset (void *__dest, int __ch, size_t __len) throw () {
if (__builtin_constant_p (__len) && __len == 0)
__warn_memset_zero_len ();
}
intent. Of course the testcase is simply reduced too much here and
the C++ FE puts the unreachable() in because of the missing return.
