[Bug rtl-optimization/124624] Missed optimization on arrangement of code blocks/branches (-Os mode)

[Explorer09 at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=124624

--- Comment #5 from Kang-Che Sung <Explorer09 at gmail dot com> ---
(In reply to Shivam Gupta from comment #4)
> Is this issue fixed on trunk, from godbolt - https://godbolt.org/z/3hjTTzb75:
> 
> trunk - 
> .L5:
>         movl    %esi, %ecx
>         movl    $-1, %edx
>         shrl    $16, %ecx
>         cmovne  %edx, %eax
>         ret
> 
> 15.1 -
> .L6:
>         or      eax, -1
>         ret
> .L5:
>         mov     ecx, esi
>         shr     ecx, 16
>         je      .L4
>         jmp     .L6
> ...
> .L4:
>         ret

I would argue it's no. The code looks worse in GCC 16.1.

For a fairer comparison, you might use -Oz instead of -Os. -Oz *should* disable
 some CMOV transformations that are unhelpful in code size reduction, and in
this case GCC's use of `mov edx, -1 /*...*/ cmovne eax, edx` increases the code
size.

See this link:
<https://godbolt.org/z/1vsfxnPj4>