[Bug rtl-optimization/107051] redundant loads when copying a union

["absoler at smail dot nju.edu.cn" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=107051

--- Comment #2 from absoler at smail dot nju.edu.cn ---
(In reply to Richard Biener from comment #1)
> With -O2 I see
> 
> func_1:
> .LFB0:
>         .cfi_startproc
>         movl    e(%rip), %eax
>         testl   %eax, %eax
>         je      .L2
> .L3:
>         jmp     .L3
>         .p2align 4,,10
>         .p2align 3
> .L2:
>         movq    g_284+8(%rip), %rax
>         movq    %rax, g_284(%rip)
>         ret
> 
> note that with -O1 we retain
> 
>   c = g_284[1];
>   c$f0_3 = g_284[1].f0;
>   c.f0 = c$f0_3;
>   g_284[0] = c;
> 
> after GIMPLE optimization which possibly explains this compared to
> 
>   c = g_284[1];
>   g_284[0] = c;
> 
> with -O2.

for gcc-13.2.0 -O2, it seems still forget to remove the load for this reduced
case:
```
union U0 {
    short f1;
    int f2;
};
union U0 g1, g2;

volatile int flag;
void func_1() {
    union U0 d[1] = {{.f1 = 1}};
    for (; flag;)
        ;
    d[0] = g2;
    g1 = d[0];
}
```
```
func_1:
.LFB0:
        .cfi_startproc
        .p2align 4,,10
        .p2align 3
.L2:
        movl    flag(%rip), %eax
        testl   %eax, %eax
        jne     .L2
        movl    g2(%rip), %eax
        movw    g2(%rip), %ax
        movl    %eax, g1(%rip)
        ret
        .cfi_endproc
```