Take the following code:
struct s
{
short t;
short y1;
};
void f(struct s *x, struct s *y)
{
struct s t = *x;
*y = t;
}
In 3.4.0, we produced:
_f:
lwz r0,0(r3)
stw r0,0(r4)
blr
but in 4.0.0 and above:
_f:
lhz r2,2(r3)
lha r0,0(r3)
sth r2,2(r4)
sth r0,0(r4)
blr
I am starting to think we should use VIEW_CNVERT_EXPR in SRA and store it into a type which is the
same as the mode if the mode is not BKL_MDE (RTL mode by the way) and we don't access the
temporary variable's fields. That would fix PR 22156 also. Maybe I can look into doing this.

Additional Comments From pinskia at gcc dot gnu dot org 2005-06-23 19:49
I have a fix, it does what I recommended.

Additional Comments From pinskia at gcc dot gnu dot org 2005-07-04 10:26
I am no longer working on this, this is a much harder problem than I sugested. This is basically the
same as PR 18268.

Additional Comments From rguenth at gcc dot gnu dot org 2005-09-15 09:46
Can you elaborate on why the VIEW_CNVERT_EXPR experiment failed and perhaps
attach the patch you tried? does using VIEW_CNVERT_EXPR just expose that
we don't do structure propagation?

Additional Comments From pinskia at gcc dot gnu dot org 2005-09-15 13:59
(In reply to comment #3)
Can you elaborate on why the VIEW_CNVERT_EXPR experiment failed and perhaps
attach the patch you tried? does using VIEW_CNVERT_EXPR just expose that
we don't do structure propagation?

VCE failed because it causes some missed optimizations and really just exposed the fact we don't have
structure copy prop.

Additional Comments From pinskia at gcc dot gnu dot org 2005-09-15 14:27
I am just going to mark this one as a dup of bug 22156, because it is the same problem, just different
testcases.
This bug has been marked as a duplicate of 22156