https://gcc.gnu.org/bugzilla/show_bug.cgi?id=57546
Andrew Pinski <pinskia at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Keywords| |rejects-valid, wrong-code
Ever confirmed|0 |1
Summary|templated gnu multiversion |gnu multiversion function
|function pointer returns |pointer returns wrong value
|wrong value |or gets rejected
Last reconfirmed| |2021-08-12
Status|UNCONFIRMED |NEW
--- Comment #3 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Confirmed, starting in GCC 6.4 (and 7+), GCC rejects it (even a non-template
version):
<source>: In constructor 'constexpr Foo<int>::Foo()':
<source>:3:7: error: cannot convert 'int (*)(Foo<int>*)' to 'int
(Foo<int>::*)()' in initialization
class Foo
^~~
<source>: In function 'int main()':
<source>:19:12: note: synthesized method 'constexpr Foo<int>::Foo()' first
required here
Foo<int> f;
^
Which is totally bogus.
note even the non-templated version is wrong.
Here is that:
class Foo
{
public:
/* Default version of foo. */
__attribute__ ((target("default")))
int foo () { return 0; }
/* corei7 version of foo. */
__attribute__ ((target("arch=corei7")))
int foo () { return 0; }
};
#include <stdio.h>
int main ()
{
Foo f;
int i = f.foo();
int (Foo::*p1)() = &Foo::foo;
int j = (f.*p1)();
if (i==j && j==0) {
printf("PASSED\n");
return 0;
} else {
printf("FAILED i:%d j:0x%x\n", i, j);
return -1;
}
}
----- CUT ----
Note for the above testcase GCC 6.4.0 and 7.1.0 both actually pass at -O0 but
ICEs at -O2 and they actually do the right thing for -O0:
mov QWORD PTR [rbp-32], OFFSET FLAT:_ZN3Foo3fooEv.ifunc(Foo*)
