Re: [PATCH v9 0/2] MR821: Replace obsolete RtlUniform tests that no longer work (currently skipped) since Windows Vista; update RtlUniform impl.

[Jinoh Kang (＠iamahuman)]

Jinoh Kang (@iamahuman) commented about dlls/ntdll/rtl.c:
> +     * the 32-bit addition above to handle.
> +     *
> +     * This produces an integer in the range [0x7fffffe2, 0xfffffffe].
> +     */
> +    result = 0x7fffffffu * 2 - result;
> +
> +    /* Perform the 2nd round of the modulus calculation.
> +     * This produces an integer in the range [0, 0x7fffffff].
> +     */
> +    result = (result & 0x7fffffffu) + (result >> 31);
> +
> +    /* If result is 0x7fffffff, set it to 0.  We avoid branching here so that
> +     * RtlUniform runs in near-constant time.  This also avoids unexpected
> +     * performance hit due to polluted branch target buffer.
> +     */
> +    result &= -((0x7fffffffu * 2 - result) >> 31);
> If we assume that the right shift operator is always defined as arithmetic shift, we can optimize it further:

```suggestion:-0+0
    result &= (LONG)(0x7fffffffu * 2 - result) >> 31;
```

That said, GCC (https://godbolt.org/z/65od5ErMo) does this optimization already. Meanwhile clang (https://godbolt.org/z/46x8a8r9v) seems to recognize the trick and replace it with `inc + cmovns` (since `0x7fffffff + 1` is `0x80000000` which sets SF).

-- 
https://gitlab.winehq.org/wine/wine/-/merge_requests/821#note_23433