https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69504
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Keywords| |accepts-invalid,
| |ice-on-invalid-code
Status|UNCONFIRMED |NEW
Last reconfirmed| |2016-01-27
Component|tree-optimization |c
Ever confirmed|0 |1
--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
This is because you use the vector extension of indexing a vector with [] but
with variable indexes the FE represents this in a way that takes the address
of the vector variable which isn't possible in this case - it's a register.
printf ((const char * restrict) "%d\n", (int) *((unsigned char *) &u8x16 +
(sizetype) i));
so I fear the solution here is to reject the code with an error or force
the frontend to special-case register vars by going through a temporary,
that is, use ({ u8x16_t tem = u8x16; tem[i]; })
Frontend issue. If you do sth like
void foo (u8x16_t *);
int main(void) {
register u8x16_t u8x16 asm ("xmm0");
foo (&u8x16);
}
the FEs will reject it with
error: address of register variable ‘u8x16’ requested
so I think rejecting this code is a reasonable thing. Using intrinsics if you
use processor specific registers is a reasonable thing to expect from you here.
