[Bug tree-optimization/119085] [12/13/14/15 regression] tree-sra generates wrong code for unions containing structs with holes (Emacs crashes with -O3) since r10-6321-g636e80eea24b78

[luangruo at yahoo dot com via Gcc-bugs]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=119085

--- Comment #7 from Po Lu <luangruo at yahoo dot com> ---
"pipcet at protonmail dot com" <gcc-bugzi...@gcc.gnu.org> writes:

> I think the problem is that the representative access chosen in
> sort_and_splice_var_accesses won't be the union copy (which would copy the
> whole storage) but the copy of the member struct (which doesn't).
>
> Making that function choose the union copy (if there is one, and we can detect
> it) fixes the problem, without giving up on scalarization for the entire 
> access
> group, but I don't know how to do it properly, and what to do if there is no
> union copy to be chosen.

The same issue affects two struct accesses of different layouts but of
the same size.  E.g., with this code:

#include "reduction.h"

struct selection_input_event kbd_buffer[KBD_BUFFER_SIZE];
struct selection_input_event *kbd_store_ptr;

void
kbd_buffer_store_buffered_event (union buffered_input_event *event,
                                 struct input_event *hold_quit)
{
  if (event->kind == ASCII_KEYSTROKE_EVENT)
    *hold_quit = event->ie;
  else
    *kbd_store_ptr = event->sie;
}

void
init_keyboard (void)
{
  kbd_store_ptr = kbd_buffer;  
}

struct selection_input_event *
get_kbd_fetch_ptr (void)
{
  return kbd_buffer;
}

If struct input_event and struct selection_input_event are padded to the
same size (on x86_64-pc-linux-gnu).

struct input_event
{
  ENUM_BF (event_kind) kind;
  unsigned code;
  unsigned modifiers;
  long device;
};
struct selection_input_event
{
  ENUM_BF (event_kind) kind;
  int *dpyinfo;
  long pad;
};