Re: [PATCH v2 0/1] MR10471: Draft: unwind: fix some invalid `reinterpret_cast`s

[Alfred Agrell (＠Alcaro)]

On Sat Mar 28 00:01:53 2026 +0000, Safocl Stollmannovic wrote:
> > None of them deem it UB. https://godbolt.org/z/17hrvMrv4
> bar(foo*, unsigned int):
>         xor     eax, eax
>         test    esi, esi
>         mov     esi, esi
>         setne   al
>         sal     eax, 16
>         add     eax, DWORD PTR [rdi+4+rsi*4]
>         ret
>
> As far as I understand, only one branch is used here. So, the
> manifestation of UB is obvious.
> UB is not a thing at asm level. The closest you'll get is a race condition, and even that will only corrupt the affected memory address, nothing else.

If the compilers deemed n!=0 to be UB and optimized accordingly, then the resulting asm would be `mov eax, [rdi+4]; ret` and nothing else.

I have no idea what you think that second branch would do. Do you want buffer overflows to jump to `abort` or [something](https://godbolt.org/z/3dnab7d9q)?

-- 
https://gitlab.winehq.org/wine/wine/-/merge_requests/10471#note_134188