[Bug c++/63157] may_alias doesn't work as expected in template nested types

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63157

--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> ---
(In reply to haynberg from comment #2)
> I recently read your old PR about placement new -
> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=29286.
> 
> Is using placement new also a means to prevent strict-aliasing optimizations
> here?  I don’t think so.  Since the following also aborts.

No, placement new constructs a new object so the contents you read
below are undefined.

> $ cat t.cpp
> #include <memory>
> #include <stdlib.h>
> 
> struct msg {
>    long seq_no;
> };
> 
> void check(char *p)
> {
>    short *a = new (p) short;
>    *a = 5;
> 
>    msg *b = new (p) msg;
>    b->seq_no = 6;
> 
>    a = new (p) short;
>    if (*a == 5)
>      abort();
> }
> 
> int main()
> {
>    msg m[1];
>    check((char*) m);
> }
> $ g++ -O3 -fno-strict-aliasing t.cpp && a.out
> $ g++ -O3 t.cpp && a.out
> Aborted